CN118072163A - Neural network-based method and system for detecting illegal occupation of territorial cultivated land - Google Patents

Abstract

The invention relates to the technical field of computer vision, and particularly discloses a method and a system for detecting illegal occupation of a cultivated land in China with high precision based on a neural network. Firstly, the SSD is integrally changed into a backbone network by using the adjusted ResNet-50; secondly, an embedded feature fusion module (R50-FFM) fuses high and low layers, and a Shortcut module (additional feature layer) is added to generate a predicted feature pyramid; then, adding a Joint Weight Adjustment Module (JWAM) to enhance useful target information; finally, the aspect ratio of the candidate frames is optimized by adopting a K-means clustering algorithm, so that the candidate frames are more suitable for homemade data sets. Experimental results show that the improved SSD target detection neural network model can improve detection accuracy and effectively improve the detection effect of small targets.

Description

High-precision detection method and system for illegal occupation of cultivated land based on neural network

Technical Field

The present invention relates to the field of computer vision technology, and in particular to a high-precision detection method and system for illegal occupation of national arable land based on a neural network.

Background technique

In the early days, land and resources monitoring consumed a lot of manpower and material resources, and its timeliness and accuracy were low. Later, with the rapid development of science and technology, more and faster and more efficient technologies were derived. Satellite remote sensing technology is rich in information and has a wide monitoring range. It mostly uses change and pattern analysis detection, but its image resolution is insufficient, resulting in low detection accuracy, poor timeliness, and susceptibility to orbit and cloud environment restrictions; UAV aerial photography technology is more suitable for the monitoring of small and medium-sized targets in cultivated land, but it requires a lot of manpower and material resources, and cannot be monitored in real time, and has poor stability in long-term monitoring of a certain place.

Summary of the invention

The present invention provides a high-precision detection method and system for illegal occupation of national arable land based on a neural network, and solves the technical problem of how to achieve high-precision real-time intelligent detection of illegal occupation of national arable land.

In order to solve the above technical problems, the present invention provides a high-precision detection method for illegal occupation of national land and cultivated land based on a neural network, comprising the steps of:

The SSD target detection neural network model is improved to obtain an improved SSD target detection neural network model;

Collect pictures of illegal occupation of cultivated land within the scope of the country, classify the types of illegal occupation, and build a data set;

The data set is used to train and learn the national land and cultivated land illegal occupation detection model to obtain the national land and cultivated land illegal occupation detection model;

The national land and cultivated land illegal occupation detection model is used to perform illegal occupation detection on the national land and cultivated land image to obtain the detection result.

Furthermore, the SSD target detection neural network model is improved as follows:

The feature extraction network VGG-16 of the SSD target detection neural network model is replaced with ResNet-50, and a feature fusion module is embedded in ResNet-50 to fuse high- and low-level information;

Adding an extra feature layer after ResNet-50 generates a prediction feature pyramid;

A joint weight adjustment module is added before the output feature maps of the first three layers of the prediction feature pyramid to perform joint weight adjustment;

The K-means clustering algorithm is used to optimize the aspect ratio of the candidate box to make it suitable for the dataset.

Furthermore, the improved SSD target detection neural network model includes a feature extraction module, a prediction feature pyramid module and a prediction feature module;

The feature extraction module uses the first 4 convolutional layers of ResNet-50 to extract features, and uses the feature fusion module to fuse the output features of the last 3 layers of the 4 convolutional layers to obtain a fused feature map;

The prediction feature pyramid module uses the fused feature map as the first prediction feature layer, and adds an additional feature layer on each subsequent prediction feature layer of the first prediction feature layer to generate the remaining five prediction feature layers, thereby obtaining a prediction feature pyramid consisting of six prediction feature layers; a joint weight adjustment module is added before the output feature maps of the first three layers of the prediction feature pyramid to perform joint weight adjustment;

The prediction feature module generates a set of candidate boxes by sliding windows from each feature map of the prediction feature pyramid, and uses the K-means clustering algorithm to determine the aspect ratio of the candidate box so that it is suitable for the data set. The number and proportion of the candidate boxes are set according to requirements; then, each candidate box is classified to determine whether it contains a target and its category is determined according to the category score; then, bounding box regression prediction is performed on the candidate box containing the target; finally, the non-maximum suppression algorithm is used to eliminate redundant boxes to obtain the position and category of the final predicted target.

Furthermore, the feature fusion module performs 1×1 convolution and bilinear interpolation operations on the feature maps output by the second, third and fourth convolutional layers conv2_3, conv3_4 and conv4_6 of ResNet-50 to adjust each feature map to the same scale, and then performs channel superposition to obtain fused features; the second, third and fourth convolutional layers conv2_3, conv3_4 and conv4_6 of ResNet-50 are composed of 3, 4 and 6 Bottlenecks respectively, each Bottleneck is connected by three layers of convolutions Conv1×1, 3×3 and 1×1, and the input and output are added by downsampling the identity mapping branch and Conv1×1, and the activation function of each Bottleneck adopts Leaky-Relu and adds BN normalization operation; the maximum pooling is added after the first convolutional layer of ResNet-50;

Each additional feature layer adopts a structure identical to the Bottleneck.

Furthermore, the joint weight adjustment module performs convolution on the input feature map to obtain a feature map F; then the channel weight adjustment module is used to adjust the proportion of each channel of the feature map F to obtain a feature map _Mc (F); then the feature map _Mc (F) is multiplied by the feature map F to obtain a feature map F'; then the position weight adjustment module is used to adjust the proportion of each pixel point of the feature map F' to obtain a feature map _Ms (F'); finally, the feature map F' is multiplied by the feature map _Ms (F') to obtain a feature map F".

Furthermore, the operation taken by the channel weight adjustment module is expressed by the formula:

_Mc (F)＝σ{MLP[AvgPool(F)]+MLP[MaxPool(F)]}

Among them, AvgPool() represents average pooling, MaxPool() represents maximum pooling, MLP[] represents multi-layer perceptron, and σ{} represents sigmoid function;

The operation taken by the position weight adjustment module is expressed by the formula:

_Ms (F') = σ{f7 ^×7 [AvgPool(F');MaxPool(F')]}

Among them, f ^7×7 [] indicates that convolution is performed using a 7×7 convolution kernel.

Furthermore, the prediction feature module uses a K-means clustering algorithm to determine the aspect ratio of the candidate box, specifically including the steps of:

1) Obtain the four coordinate values [X _min , Y _min , X _max , Y _max ] of the real box in the data set, where X _min and Y _min represent the horizontal and vertical coordinates of the lower left corner of the real box, respectively, and X _max and Y _max represent the horizontal and vertical coordinates of the upper right corner of the real box, respectively, and calculate the corresponding aspect ratio based on the four coordinate values;

2) Initialization, pre-set k cluster centers;

3) Calculate the intersection-and-union ratio of the true box and the cluster center in turn, and assign the true box to the cluster with the minimum intersection-and-union ratio;

4) After all the real boxes are assigned, the location of the cluster center is recalculated;

5) Determine whether the position of the cluster center has changed. If so, return to step 3). If not, obtain the latest K cluster centers.

6) Determine the aspect ratio of the candidate box based on the determined K cluster centers.

Furthermore, the aspect ratios of the candidate boxes include six ratios: 1, 2/3, 3/2, 4/5, 6/5, and 9/5.

The present invention also provides an improved SSD national land and farmland illegal occupation detection system using the above method, the key of which is: including a model construction module, a data set construction module, a model training module and a model application module;

The model building module is used to improve the SSD target detection neural network model to obtain an improved SSD target detection neural network model;

The data set construction module is used to collect pictures of illegal occupation behaviors within the scope of national arable land, classify illegal occupation types, and construct a data set;

The model training module is used to train and learn the national land and cultivated land illegal occupation detection model using the data set to obtain the national land and cultivated land illegal occupation detection model;

The model application module is used to use the national land and cultivated land illegal occupation detection model to perform illegal occupation detection on the national land and cultivated land image to obtain the detection result.

The present invention also provides a computer-readable storage medium, the key of which is that a computer application is stored thereon, and the program is executed by a processor to implement the above-mentioned high-precision detection method for illegal occupation of national land based on neural network.

The high-precision detection method and system for illegal occupation of cultivated land based on neural network provided by the present invention, aiming at various drawbacks of traditional technology, uses the combination of high-definition video and deep learning, relies on the video data collected in real time by high-definition camera, and uses the improved SSD target detection algorithm to quickly complete the high-precision real-time intelligent detection of illegal occupation of cultivated land. First, the SSD overall uses the adjusted ResNet-50 as the backbone network; secondly, the feature fusion module (R50-FFM) is embedded to fuse the high and low layers, and the Shortcut module (extra feature layer) is added to generate the prediction feature pyramid; then, the joint weight adjustment module (JWAM) is added to enhance the useful target information; finally, the K-means clustering algorithm is used to optimize the aspect ratio of the candidate box to make it more suitable for self-made data sets. Experimental results show that the improved SSD target detection neural network model can improve the detection accuracy and effectively improve the detection effect of small targets. The final model mAP value is 91.24%, which is 5.74% higher than SSD, and the detection rate is 40.3fps, which can be used as a reference for land and resources monitoring.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a framework diagram of an SSD algorithm provided by an embodiment of the present invention;

FIG2 is a framework diagram of an improved SSD algorithm provided by an embodiment of the present invention;

FIG3 is a schematic diagram of a feature fusion module provided in an embodiment of the present invention;

FIG4 is a structural diagram of R50-FFM provided in an embodiment of the present invention;

FIG5 is a diagram showing the internal structure of an additional layer provided in an embodiment of the present invention;

FIG6 is a structural diagram of a joint weight adjustment module provided in an embodiment of the present invention;

7 is a structural diagram of a channel weight adjustment module provided in an embodiment of the present invention;

FIG8 is a structural diagram of a position weight adjustment module provided in an embodiment of the present invention;

FIG9 is a flow chart of a K-means clustering algorithm provided in an embodiment of the present invention;

FIG10 is a visualization diagram of cluster centers provided by an embodiment of the present invention;

FIG11 is an example diagram of an SSD algorithm detection process provided by an embodiment of the present invention;

FIG12 is a comparison diagram of the visual detection effect before and after the improvement provided by an embodiment of the present invention, wherein (a) corresponds to before the improvement, and (b) corresponds to after the improvement;

FIG13 is a comparison diagram of backbone network accuracy and loss provided by an embodiment of the present invention;

FIG. 14 is a comparison chart of ablation experiment accuracy trends provided by an embodiment of the present invention.

Detailed ways

The following specifically illustrates the implementation mode of the present invention in conjunction with the accompanying drawings. The embodiments are provided for illustrative purposes only and are not to be construed as limitations of the present invention. The accompanying drawings are provided for reference and illustration only and do not constitute limitations on the scope of patent protection of the present invention, because many changes may be made to the present invention without departing from the spirit and scope of the present invention.

Nowadays, it is common to erect towers in cultivated land. Using images obtained by cameras mounted on towers for image research has the advantages of good real-time performance, high clarity, strong stability, convenience and flexibility. Deep learning, as a hot branch in the field of AI, has developed rapidly in computer vision and other aspects, and has also greatly promoted the progress of AI.

In view of the various drawbacks of traditional technologies, this paper combines high-definition video with deep learning, relying on the video data collected in real time by high-definition cameras mounted on communication towers built by China Tower Corporation, and uses the SSD target detection algorithm to quickly complete the real-time intelligent detection of illegal occupation of cultivated land, and further improves the SSD algorithm to achieve satisfactory results. The realization of intelligent rural land planning greatly reduces the cost of manual inspections and improves government work efficiency, which is of great significance to the construction and development of future smart cities and smart land.

Specifically, the high-precision detection method for illegal occupation of cultivated land based on a neural network provided in an embodiment of the present invention comprises the following steps:

S1. Improve the SSD target detection neural network model to obtain an improved SSD target detection neural network model;

S2. Collect pictures of illegal occupation of cultivated land within the scope of the country, classify the types of illegal occupation, and construct a data set;

S3, using the data set to train and learn the national land and farmland illegal occupation detection model to obtain the national land and farmland illegal occupation detection model;

S4. Use the national land and farmland illegal occupation detection model to detect illegal occupation of national land and farmland images to obtain detection results.

(1) Step S1: Improving the SSD target detection neural network model

The SSD target detection neural network model or SSD algorithm (Single Shot MultiBox Detector, SSD) is a classic single-stage target detection algorithm proposed by Liu Wei in ECCV in 2016. It uses the anchor mechanism of Faster-RCNN. Its algorithm framework is shown in Figure 1. It uses VGG-16 as a feature extraction network, discards its fully connected layer and converts it into a convolutional layer, and adds a series of convolutional layers to obtain prediction results. Feature maps of different scales are used for detection. Large-scale feature maps are used to detect small objects, and small-scale feature maps are used to detect large objects.

The input image size of SSD300 is 300×300. First, the input image features are extracted using the pre-trained feature extraction network to obtain feature maps of different scales in 6 prediction feature layers, namely Conv4_3, Conv7, Conv8_2, Conv9_2, Conv10_2, and Conv11_2. Secondly, a series of candidate boxes are generated for each feature map by sliding windows, and the number and proportion of the boxes are set manually. Then, each candidate box is classified to determine whether it contains a target and its category is determined according to the category score. Furthermore, bounding box regression prediction is performed on the candidate boxes containing the target to locate the target more accurately. Finally, the non-maximum suppression (NMS) algorithm is used to remove redundant boxes to obtain the position and category of the final predicted target.

In order to improve the model detection accuracy, the present invention improves the SSD target detection neural network model, and the architecture of the obtained improved SSD target detection neural network model is shown in Figure 2. The present invention adopts ResNet-50 with better feature extraction ability, stronger network representation ability and relatively low computational complexity as the SSD backbone network improvement model framework, and deletes and adds a series of additional layers to construct the entire SSD network. The introduction of the residual module ensures that as the network deepens, the high-order semantic features of the image will not disappear, avoiding the problem of gradient disappearance and explosion. ResNet-50 is organically composed of a series of bottleneck layers (Bottleneck). Each Bottleneck is composed of three layers of convolutions: Conv1×1, 3×3, and 1×1, and the input and output addition is achieved through the identity mapping branch downsampling and Conv1×1.

Remove conv5_x and subsequent structures of ResNet-50, retain the first four convolutional layers, add maximum pooling after the first layer Conv7×7, and the last three layers consist of 3, 4, and 6 Bottlenecks respectively. Replace the activation function Relu with Leaky-Relu and add BN (Batch Normal) normalization operation.

As the convolutional layer network deepens, the semantic features extracted by the network become stronger and stronger, and the high-resolution shallow feature map has extremely strong positioning information. The pixel ratio of small targets is small, and its semantic information gradually becomes blurred after multiple downsampling. SSD treats the 6 layers of feature maps of different degrees equally, and cannot fully utilize the respective advantages of local detail features (positioning) and global semantic features (recognition). Therefore, this example uses a feature fusion module (Feature Fusion Module, R50-FFM) different from FPN to combine shallow detail features with high-level semantic features, so that each feature map has high positioning and recognition information, enhancing the detection effect of small targets.

As shown in Figure 3, this example uses channel superposition (Concat) for feature fusion. This method has the characteristics of integrating semantic information of different levels and spaces. Specifically, each layer of feature map is subjected to 1×1 convolution and bilinear interpolation operations so that the corresponding elements of the high-level and low-level feature maps can be channel-superimposed for feature fusion.

The feature fusion function φ _f is used to fuse the feature maps Xi that have been adjusted to a uniform scale by the scale adjustment function _{ζ i into} X _f , so that the feature pyramid function φ _p can be used to generate a new feature pyramid X' _p based on the fusion layer X _f , and then the function φ _c,l is used to predict the target location (loc) and classification (class) from the feature pyramid. The process can be expressed as:

Among them, i∈(2-3,3-4,4-6) corresponds to the last three convolutional layers in the first four layers of ResNet-50, and the function φ _p is used to generate 6 feature pyramids (feature maps) of different scales, and a total of 8732 bounding boxes are predicted. When predicting the bounding boxes, the function φ _c,l and NMS methods are used to obtain the final target position and category from the 8732 bounding boxes. ∪(X' _p ) represents the operation of generating bounding boxes for generating a new feature pyramid X' _p .

In this example, the feature fusion module is embedded in the SSD backbone network ResNet-50, and an additional feature layer is added afterwards to generate a feature fusion pyramid. The three convolutional layers conv2_3, conv3_4, and conv4_6 of ResNet-50 are adjusted to the same scale for Concat connection, thereby realizing multi-scale feature fusion of high and low layers. Specifically, as shown in Figure 4, the above three layers generate feature maps of sizes 75×75×256, 38×38×512, and 19×19×1024 respectively through convolution, and then unified to 38×38×1024. The feature map obtained by BN layer processing is used as the first prediction feature layer Feature Map1 of SSD.

As shown in Figure 5, the internal structure of the additional added layer, this example also adds 5 shortcut modules (Bottleneck operation) in sequence as additional added layers to generate the remaining 5 SSD prediction feature layers Feature Map2 to Feature Map6, with scales of 19×19×1024, 10×10×512, 5×5×512, 3×3×256, and 1×1×256. Feature Map1 to Feature Map6 together form a feature fusion pyramid as the feature map of the SSD prediction feature layer.

The Joint Weight Adjustment Module (JWAM) can be used as a "small plug-in" in convolutional neural networks. It is independent of other network models and is widely used in target detection, image recognition and other fields. Inspired by how humans process visual image information, it can focus on useful information with high weights, ignore irrelevant information with low weights, and continuously adjust the weights so that it can obtain more and more important information in different situations, to ensure the usefulness of the convolutional layer space and channels, and make positioning and classification more accurate.

In this example, a joint weight adjustment module is added before the output feature map of the first three layers of the prediction feature pyramid to perform joint weight adjustment. The structure of the joint weight adjustment module is shown in Figure 6. After the joint weight adjustment module performs convolution on the input feature map, a feature map F is obtained; then the channel weight adjustment module is used to adjust the weight of the proportion of each channel of the feature map F to obtain the feature map _Mc (F); then the feature map _Mc (F) is multiplied with the feature map F to obtain the feature map F'; then the position weight adjustment module is used to adjust the weight of the proportion of each pixel point of the feature map F' to obtain the feature map _Ms (F'); finally, the feature map F' is multiplied with the feature map _Ms (F') to obtain the feature map F".

The Channel Weight Adjustment Module (CWAM) is used to focus on the target channel information and adjust the weights of the proportions of each channel in the image. As shown in Figure 7, the module performs spatial maximum pooling and mean pooling on each channel of the feature map F to obtain the feature map . Arrange the maximum and mean pooling values in channel order to obtain two result vectors. Then send them to the two fully connected layers (multi-layer perceptron MLP) for dimensionality reduction and dimensionality increase operations. Finally, add the output feature results and activate them with the σ function (sigmoid function) to generate the channel adjustment vector M _C , that is, the weight of each channel. The process is expressed as:

Among them, _W0 and _W1 represent the weights of the two fully connected layers.

As for the Position Weight Adjustment Module (PWAM), this module is used to focus on the target position information and adjust the weight of each pixel ratio in the image. As shown in Figure 8, this module performs maximum and mean pooling at the same position of each channel of the feature map F' to obtain the feature map , and then merge the two pooling results to obtain the feature map of (W, H, 2) (W is the width, H is the height, and 2 is the number of channels). Finally, use a 7×7 convolution kernel to activate the convolved feature map through the σ function to obtain the position adjustment vector _MS . The process is expressed as follows:

The output of the channel module _Mc and F are multiplied by points to obtain F' as the input of the position module to achieve the connection between the two, and the whole constitutes a joint weight adjustment module. The input feature map F is adjusted by weight to obtain the final output feature map F', which can be expressed as:

Since the information contained in the feature map becomes increasingly blurred when the resolution is less than 10×10, JWAM is inserted into the first three feature maps to focus on the target information.

Candidate box generation requires setting the scaling ratio _sk and aspect ratio a _r .

Scaling ratio _sk , the ratio of the candidate box size to the input image. _sk is set to follow a linear increase rule:

Since the first layer is set independently, m is set to 5 and s _min /2 = 0.1. s _min and s _max are the upper and lower limits of the ratio, which are set to 0.2 and 0.9. The scales of the candidate boxes of the 2nd to 6th layers increase linearly according to formula (5).

Aspect ratio a _r {1, 1*, 2, 1/2, 3, 1/3}, a total of 6 ratios are used to generate candidate boxes. When _{a r} = 1, a square bounding box with a scaling ratio of _sk is generated; a _r = 1*, a square bounding box with a scaling ratio of sk is generated. The actual width and height of each candidate box:

Candidate box and ground truth (GT) matching principle: Calculate the IOU value of each candidate box and GT in turn. If it is greater than a certain threshold (usually 0.5), the candidate box matches GT and will be responsible for predicting GT, which is a positive sample; otherwise, it is a negative sample. To avoid imbalance between positive and negative samples, SSD samples negative samples and selects them by confidence error, controlling the ratio of positive and negative samples to 1:3.

The candidate boxes pre-set by the SSD algorithm were originally for the Pascal VOC2007 dataset, and the scale setting is more inclined to large and medium targets. The predicted boxes are decoded from the candidate boxes and cannot meet the detection requirements for small targets. Therefore, it is necessary to further adjust the width and height ratio of the candidate boxes for the self-built dataset to increase the matching probability between the target and the candidate box, and further improve the detection accuracy.

This paper uses the K-means clustering algorithm to optimize the aspect ratio of the candidate box. The aspect ratio that is more suitable for the dataset in this paper is obtained by clustering. The specific steps include:

1) Obtain the four coordinate values [X _min , Y _min , X _max , Y _max ] of the real box in the data set, where X _min and Y _min represent the horizontal and vertical coordinates of the lower left corner of the real box, respectively, and X _max and Y _max represent the horizontal and vertical coordinates of the upper right corner of the real box, respectively, and calculate the corresponding aspect ratio based on the four coordinate values;

2) Initialization, pre-set k cluster centers;

3) Calculate the distance between the true box and the cluster center in turn, and assign the true box to the cluster with the minimum distance;

4) After all the real boxes are assigned, the location of the cluster center is recalculated;

5) Determine whether the position of the cluster center has changed. If so, return to step 3). If not, obtain the latest K cluster centers.

6) Determine the aspect ratio of the candidate box based on the determined K cluster centers.

Traditional clustering methods use Euclidean distance to measure the distance between the true box and the cluster center, but this method has a larger error when the box size is larger. The IOU value is used as a proxy for the Euclidean distance to avoid this problem.

The IOU value is measured as follows:

Among them, IOU is the intersection-over-union ratio, G is the true box, K is the cluster box, C is the cluster center, and D is the measurement index. The smaller D is, the closer the true box is to the cluster box.

By calculation, it is found that when K=9, the average IOU value is the highest, that is, the accuracy is 74.68%, and its aspect ratios are (31.875, 42.15), (49.82, 75.05), (91.63, 81), (66.5, 129), (142.65, 116.94), (102.92, 173.25), (168, 207.1), (255.9, 156.4), (256.82, 262.88), and the 9 cluster centers are shown in Figure 10. After data analysis, 1, 2/3, 3/2, 4/5, 6/5, and 9/5 are finally selected as the width-to-height ratios of the candidate boxes.

(2) Step S2: Create a dataset

The monitoring of illegal occupation of cultivated land to be realized in this paper is based on the ROI (region of interest) delineated by the Bureau of Land and Resources, and does not include the illegal occupation of cultivated land that has existed before. The common illegal occupation behaviors within the scope of national cultivated land are divided into characteristic categories to construct a data set, and video capture images are obtained. 10 images are saved every hour and images without occupation targets are removed; the images are uniformly cropped to 640*550 and the Retinex defogging algorithm is used to remove interference from the image data set. Finally, 859 images of 15 categories including 6 types of engineering vehicles, 5 types of poultry, 2 types of buildings, fish ponds, and trees are selected to construct a data set.

The image dataset was expanded 6 times to 5154 images using one or more random combinations of image flipping, rotation, filtering, and color space conversion. The dataset was randomly divided into training, validation, and test sets in a ratio of 8:1:1, as shown in Table 1. The labelimg annotation tool was used to label the images to obtain the label file Annotations, which was constructed in the VOC format for training and learning of the improved SSD model.

Table 1 Dataset division

(3) Step S3: Model training

The experiment in this paper is based on the Windows platform, the CPU is AMD Ryzen 7 4800H, and the graphics card is NVIDIIA GTX1650ti. The programming language is Python-3.9. The compilation environment is torch-1.10, torchvision-0.11.0, and cuda-10.2. During network training, the weights of the improved SSD algorithm are initialized using the weights obtained by training the traditional SSD on the VOC 2007 dataset. The training parameter settings are shown in Table 2.

Table 2 Training parameter configuration

This example uses the mAP (mean Average Precision) value to evaluate the model. It comprehensively considers the category and positioning and is an important indicator for evaluating model performance in the field of object detection. In the SSD algorithm, the relationship between the target box and the generated prediction box is defined as follows.

TP: True positive, actually the target and the IOU with the generated prediction box is greater than the threshold, the target is predicted correctly; TN: True negative, actually the background and no prediction box is generated, the background is predicted correctly; FP: False positive, the background is mistakenly predicted as the target, which is a false detection of the target; FN: False negative, the target is mistakenly predicted as the background, which is a missed detection of the target.

Precision (P), the precision rate, indicates the proportion of samples predicted as positive examples that are correctly predicted:

Precision (P), the precision rate, indicates the proportion of samples predicted as positive examples that are correctly predicted:

Recall (R), the recall rate, indicates the proportion of samples that are actually positive examples that are predicted correctly:

AP (Average Precision) is used to describe the average precision of a single type, which means the integration of P and R on 0-1.

(4) Step S4: Model application

The national land and farmland illegal occupation detection model is used to detect illegal occupation of national land and farmland images generated in real time to obtain detection results.

In order to facilitate the application of the above method, this embodiment also provides an improved SSD national land and farmland illegal occupation detection system, including a model construction module, a data set construction module, a model training module and a model application module;

The model building module is used to improve the SSD target detection neural network model to obtain an improved SSD target detection neural network model;

The dataset construction module is used to collect pictures of illegal occupation behaviors within the scope of national arable land, classify the types of illegal occupation, and construct a dataset;

The model training module is used to train and learn the national land and farmland illegal occupation detection model using the data set to obtain the national land and farmland illegal occupation detection model;

The model application module is used to use the national land and farmland illegal occupation detection model to detect illegal occupation of national land and farmland images and obtain detection results.

This embodiment also provides a computer-readable storage medium on which a computer application is stored. The program is executed by a processor to implement a high-precision detection method for illegal occupation of national land based on a neural network.

(5) Experiment

The experimental data set setting and model training are consistent with the above.

1. Comparison of detection effects before and after improvement

Figure 11 shows some feature maps generated by the SSD detection image learning process. SSD finally locates the target and outputs the category with the highest score. In order to evaluate the performance of the improved algorithm and the detection effect of small targets, this paper selects 4 dataset images for detection effect comparison. As shown in Figure 12, for large and medium-sized targets with a high pixel ratio, such as large engineering vehicles, there is no obvious change in the detection before and after; while for small targets, such as sheep photographed from high altitudes, which contain less semantic information, SSD has many missed detection targets and even false detections (Figure 12 (a) misdetects sheep as cows). After the improved SSD algorithm detection, its small target missed detection rate is improved, which enhances the detection effect of small targets.

2. Comparison with other target detection algorithms

In order to verify that this algorithm can improve the detection accuracy, it is compared with other advanced target detection algorithms. The evaluation indicators are mAP value and detection speed. The same self-built dataset is used. The results are shown in Table 3.

Table 3 Performance comparison of different algorithms

From the results in Table 3, we can see that compared with the two-stage algorithm Faster R-CNN, the FPS of this improved algorithm is increased by 34.08, which shows the speed advantage of the single-stage algorithm, and the accuracy is increased by 2.22%; compared with the more classic DSSD of improved SSD, the accuracy is increased by 2.64%, and the FPS is increased by 24.7; compared with the lightweight Yolov4-tiny that focuses on detection speed, the accuracy is increased by 5.06%, and the FPS is reduced by 33.2. The experiment shows that compared with other algorithms, this algorithm has certain comprehensive performance and can meet the actual requirements of real-time detection of illegal occupation of cultivated land.

3. Ablation Experiment Analysis

First, the backbone network was replaced, and the same data set was used to train 200 epochs according to the above experimental parameter configuration, and an evaluation was performed every 10 epochs. The comparison of its mAP value and training set loss curve is shown in Figure 13. It can be clearly seen that after replacing the backbone network, the mAP is improved and the loss is reduced, which proves that the ResNet-50 network has stronger feature extraction capabilities and prevents network degradation. The results are shown in Table 4. The mAP value and detection speed are increased by 1.32% and 2.2FPS respectively. This is because the pruned ResNet-50 ensures the network depth while reducing the number of parameters.

Table 4 Backbone network comparison experiment

In order to verify the effectiveness of each improved link of this algorithm, ablation comparison tests were continued on the basis of replacing the backbone network, and R50-FFM, JWAM modules and K-means optimization were gradually added to analyze the performance of each module by comparison. Self-built data sets were used uniformly, and the evaluation indicators were mAP value and detection speed. Figure 14 shows the comparison of the accuracy trend of ablation experiments, and Table 5 shows the experimental results. The R50-FFM module was embedded alone, and the mAP increased by 2.4%, proving the effectiveness of the fusion of deep semantic information and shallow detail information. Compared with a single feature map, the fused multi-scale feature map contains more useful information. The increase in network parameters caused the FPS to drop by 5.6. Based on this, the JWAM module was added, and the mAP increased by 1.41%, proving that the JWAM module makes the network focus on the target area to improve the detection accuracy, and the FPS continues to decrease; using K-means optimization, the mAP value increased slightly by 0.49%, proving the effectiveness of optimizing the candidate frame ratio, improving the matching probability of the target and the candidate frame, and the FPS increased by 1.3. More suitable candidate frames can improve the detection speed. Combining all the improvements, the final algorithm mAP value is 91.24%, which is 4.29% higher than the baseline ResNet-50, proving the effectiveness of this algorithm improvement.

Table 5 Ablation experiment results

In summary, the present invention targets the illegal occupation of cultivated land in the country, obtains videos through tower-mounted cameras, and combines the SSD target detection algorithm to monitor the designated area, thereby improving the work efficiency of law enforcement departments. The SSD algorithm is further optimized and improved to make it more suitable for land and resources monitoring, replace the backbone network, deepen the network depth while reducing network parameters, enhance the network's anti-interference ability and feature extraction ability; add a feature fusion module so that both high and low layers have strong semantics and positioning information, and enhance the small target detection performance; add a weight adjustment module to suppress irrelevant information; use K-means to optimize the candidate frame ratio, and the accuracy is further improved. The experimental results show that the model detection accuracy is increased from 85.8% to 91.24%, which fully proves the effectiveness of the algorithm. Compared with other target detection algorithms, it also has higher comprehensive performance, can meet the actual needs of the project, and is suitable for land and resources monitoring.

The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims (10)

1. A high-precision detection method for illegal occupation of cultivated land based on a neural network, characterized in that it comprises the following steps: The SSD target detection neural network model is improved to obtain an improved SSD target detection neural network model; Collect pictures of illegal occupation of cultivated land within the scope of the country, classify the types of illegal occupation, and build a data set; The data set is used to train and learn the national land and cultivated land illegal occupation detection model to obtain the national land and cultivated land illegal occupation detection model; The national land and cultivated land illegal occupation detection model is used to perform illegal occupation detection on the national land and cultivated land image to obtain the detection result. 2. According to the high-precision detection method for illegal occupation of cultivated land based on neural network in claim 1, it is characterized in that the SSD target detection neural network model is improved, specifically: The feature extraction network VGG-16 of the SSD target detection neural network model is replaced with ResNet-50, and a feature fusion module is embedded in ResNet-50 to fuse high- and low-level information; Adding an extra feature layer after ResNet-50 generates a prediction feature pyramid; A joint weight adjustment module is added before the output feature maps of the first three layers of the prediction feature pyramid to perform joint weight adjustment; The K-means clustering algorithm is used to optimize the aspect ratio of the candidate box to make it suitable for the dataset. 3. The high-precision detection method for illegal occupation of cultivated land based on neural network according to claim 2 is characterized in that: the improved SSD target detection neural network model includes a feature extraction module, a prediction feature pyramid module and a prediction feature module; The feature extraction module uses the first 4 convolutional layers of ResNet-50 to extract features, and uses the feature fusion module to fuse the output features of the last 3 layers of the 4 convolutional layers to obtain a fused feature map; The prediction feature pyramid module uses the fused feature map as the first prediction feature layer, and adds an additional feature layer on each subsequent prediction feature layer of the first prediction feature layer to generate the remaining five prediction feature layers, thereby obtaining a prediction feature pyramid consisting of six prediction feature layers; a joint weight adjustment module is added before the output feature maps of the first three layers of the prediction feature pyramid to perform joint weight adjustment; The prediction feature module generates a set of candidate boxes by sliding windows from each feature map of the prediction feature pyramid, and uses the K-means clustering algorithm to determine the aspect ratio of the candidate boxes so that they are suitable for the data set. The number and proportion of the candidate boxes are set according to requirements. Then, each candidate box is classified to determine whether it contains a target and its category is determined according to the category score. Then, bounding box regression prediction is performed on the candidate boxes containing the target. Finally, the non-maximum suppression algorithm is used to eliminate redundant boxes to obtain the position and category of the final predicted target. 4. The high-precision detection method for illegal occupation of cultivated land based on a neural network according to claim 3 is characterized by: The feature fusion module performs 1×1 convolution and bilinear interpolation operations on the feature maps output by the second, third and fourth convolutional layers conv2_3, conv3_4 and conv4_6 of ResNet-50 to adjust each feature map to the same scale, and then performs channel superposition to obtain fusion features; the second, third and fourth convolutional layers conv2_3, conv3_4 and conv4_6 of ResNet-50 are composed of 3, 4 and 6 Bottlenecks respectively, each Bottleneck is connected by three layers of convolution, Conv1×1, 3×3 and 1×1, and the input and output are added by downsampling the identical mapping branch and Conv1×1, and the activation function of each Bottleneck adopts Leaky-Relu and adds BN normalization operation; the maximum pooling is added after the first convolutional layer of ResNet-50; Each additional feature layer adopts a structure identical to the Bottleneck. 5. According to the high-precision detection method for illegal occupation of cultivated land based on neural network in claim 4, it is characterized in that: the joint weight adjustment module convolves the input feature map to obtain a feature map F; then the channel weight adjustment module is used to adjust the proportion of each channel of the feature map F to obtain a feature map _Mc (F); then the feature map _Mc (F) is multiplied with the feature map F to obtain a feature map F'; then the position weight adjustment module is used to adjust the proportion of each pixel point of the feature map F' to obtain a feature map _Ms (F'); finally, the feature map F' is multiplied with the feature map _Ms (F') to obtain a feature map F". 6. According to the high-precision detection method for illegal occupation of cultivated land based on neural network in claim 5, it is characterized in that the operation taken by the channel weight adjustment module is expressed by the formula: _Mc (F)＝σ{MLP[AvgPool(F)]+MLP[MaxPool(F)]} Among them, AvgPool() represents average pooling, MaxPool() represents maximum pooling, MLP[] represents multi-layer perceptron, and σ{} represents sigmoid function; The operation taken by the position weight adjustment module is expressed by the formula: _Ms (F') = σ{f7 ^×7 [AvgPool(F');MaxPool(F')]} Among them, f ^7×7 [] indicates that convolution is performed using a 7×7 convolution kernel. 7. The high-precision detection method for illegal occupation of cultivated land based on neural network according to claim 6 is characterized in that the prediction feature module uses K-means clustering algorithm to determine the aspect ratio of the candidate box, specifically comprising the steps of: 1) Obtain the four coordinate values [X _min , Y _min , X _max , Y _max ] of the real box in the data set, where X _min and Y _min represent the horizontal and vertical coordinates of the lower left corner of the real box, respectively, and X _max and Y _max represent the horizontal and vertical coordinates of the upper right corner of the real box, respectively, and calculate the corresponding aspect ratio based on the four coordinate values; 2) Initialization, pre-set k cluster centers; 3) Calculate the intersection-and-union ratio of the true box and the cluster center in turn, and assign the true box to the cluster with the minimum intersection-and-union ratio; 4) After all the real boxes are assigned, the location of the cluster center is recalculated; 5) Determine whether the position of the cluster center has changed. If so, return to step 3). If not, obtain the latest K cluster centers. 6) Determine the aspect ratio of the candidate box based on the determined K cluster centers. 8. According to the high-precision detection method for illegal occupation of national land based on neural network in claim 7, it is characterized in that the aspect ratio of the candidate frame includes six ratios: 1, 2/3, 3/2, 4/5, 6/5, and 9/5. 9. A system for detecting illegal occupation of cultivated land using the improved SSD of the high-precision detection method for illegal occupation of cultivated land based on a neural network as claimed in any one of claims 1 to 8, characterized in that it comprises a model building module, a data set building module, a model training module and a model application module; The model building module is used to improve the SSD target detection neural network model to obtain an improved SSD target detection neural network model; The data set construction module is used to collect pictures of illegal occupation behaviors within the scope of national arable land, classify illegal occupation types, and construct a data set; The model training module is used to train and learn the national land and cultivated land illegal occupation detection model using the data set to obtain the national land and cultivated land illegal occupation detection model; The model application module is used to use the national land and cultivated land illegal occupation detection model to perform illegal occupation detection on the national land and cultivated land image to obtain the detection result. 10. A computer-readable storage medium, characterized in that a computer application is stored thereon, and the program is executed by a processor to implement the high-precision detection method for illegal occupation of land and cultivated land based on a neural network as described in any one of claims 1 to 8.