CN111126278B - A Method of Optimizing and Accelerating Object Detection Models for Few-Category Scenes - Google Patents

Abstract

The invention relates to the technical field of computer vision, in particular to a method for optimizing and accelerating a target detection model for few types of scenes, which comprises the following steps: acquiring and labeling a picture to be detected; adding the marked pictures into a feature extraction device to perform feature extraction to obtain three groups of feature images with fixed sizes; sending the three groups of feature images with fixed sizes to a predicting device with a FocalLoss loss function for result prediction; filtering the prediction result by using a detection frame, so that only one detection frame outputs for the same object to be detected, wherein the characteristic extraction process comprises the following steps: compressing network DarkNet, alexNet, resNet, VGG, googLeNet, SENet, denseNet, adjusting the size of the picture to N x N, wherein N is a multiple of 32 between 320 and 1280, and extracting the characteristics of the picture by adopting the compressed characteristic extraction network. The invention can simultaneously have higher detection speed and higher accuracy.

Description

A Method of Optimizing and Accelerating Object Detection Models for Few-Category Scenes

technical field

The invention relates to the technical field of computer vision, in particular to a method for optimizing and accelerating an object detection model in a scene with few categories.

Background technique

The target detection task is an important application in the field of computer vision. Its main goal is to find all target objects from the image and give their classification information and location information.

Based on traditional pattern recognition methods, it usually takes tens of seconds to process a picture, while the target detection algorithm based on convolutional neural network can already complete the target detection of a high-resolution picture in hundreds or even tens of milliseconds. Work, but also has a great improvement in accuracy.

The rapid development of neural networks has made people move away from the method of manually designing features, and turned to the method of training deep convolutional neural networks by using a large number of data sets, so that the network can learn features adaptively.

Existing target detection technology solutions based on convolutional neural networks are mainly divided into the following three categories:

The first type: two-stage detector, its main feature is to determine the position information and classification information of the object in two steps. First select the region of interest. In the early RCNN and Fast-RCNN networks, the region of interest was determined by selective search. When developing to Faster-RCNN, the region generation network was invented to determine the region of interest, so that the entire The object detection network becomes an end-to-end structure. In the second stage, the convolutional neural network further corrects the coordinate position of the region of interest extracted in the first stage, and after passing all the prediction results through the non-maximum value suppression algorithm, the final result is output, and the object in the picture is obtained. Precise position and coordinates, it can be seen that this type of method is characterized by more accurate results, but its corresponding time overhead is also relatively large.

The second category: single-stage detectors, such as YOLO, SSD, its main feature is that it only needs one step to directly obtain the location information and classification information of the object, different from the two-stage detector, the network is directly in the feature map of the last layer Output the final prediction result. For each point of the last layer feature map, output the position offset relative to the preset anchor point, and directly obtain the position information of each object. Similarly, the final result also needs to pass The non-maximum suppression algorithm, since this type of method has only one stage, is much faster than the two-stage detector, but the accuracy is relatively low.

The third category: methods without anchor points. The characteristic of this type of method is that it does not need to set anchor points in advance, and directly regresses the four boundaries of each object. At present, the accuracy can reach a level equivalent to that of a two-stage detector. However, the speed is generally slow and the practicability is poor.

Contents of the invention

The present invention aims at the defects of the prior art that the detection speed is fast but the accuracy rate is low, or the accuracy rate is high but the detection speed is slow, and provides a method for optimizing and accelerating the target detection model for a few-category scene, and has a high detection rate. speed and high accuracy.

The present invention is realized by adopting the following technical solutions:

A method for optimizing and accelerating a target detection model for a scene with few categories, comprising the following steps:

Obtain and label the image to be detected;

The marked pictures are added to the feature extraction device for feature extraction, and three sets of fixed-size feature maps are obtained;

Send three sets of fixed-size feature maps to the prediction device with FocalLoss loss function for result prediction;

Filter the detection frame of the prediction result so that only one detection frame is output for the same object to be detected;

The feature extraction process includes: compressing the feature extraction network, adjusting the size of the picture to a resolution of N×N, the selectable value of N is a multiple of 32 between 320-1280, and using the compressed feature The extraction network performs feature extraction of pictures.

Preferably, the compression of the feature extraction network is to reduce the number of convolution kernels to the original X times for all convolution layers, and the optional value of X is X={2,4,6, 8}.

Preferably, the feature extraction network is one of DarkNet network, AlexNet network, ResNet network, VGG network, GoogLeNet network, SENet network or DenseNet network.

Preferably, an adaptive module is added in the FocalLoss loss function, specifically:

For each prediction result of each picture, calculate the cross entropy loss, accumulate all the losses that are actually marked as the foreground, denoted as M, and accumulate all the losses that are truly marked as the background, denoted as N;

Calculate log(N/M), record it as S, if S is less than zero, then S is set to zero, if it is the first iteration, let the value of S ₀ be S;

According to the value of S, calculate the value of α, α=1/(2 ^S +0.2);

The value of γ is calculated according to the values of S and S ₀ , and the calculation formula is γ=min(S ₀ −S,2).

Preferably, the filtering of the detection frame comprises the following steps:

Step 1: First obtain the network output detection frame coordinate information set B={b ₁ , b ₂ ...b _n }, and the confidence set is S=[s ₁ , s ₂ ...s _n }, where b _t , s _t corresponds to the same detection frame coordinate information and confidence;

Step 2: Take out the element with the highest confidence from S, and mark it as i;

Step 3: Get s _i and b _i according to subscript i, and delete s _i and b _i from sets S and B respectively;

Step 4: Traverse the remaining detection frames in the set B, and add the subscripts of all detection frames with an intersection ratio greater than 0.5 with _bi to the set T. If T is not an empty set, add bi _and s _i to In the result sets D and F, delete the elements whose subscripts are in the set T from the S and B sets at the same time, if T is an empty set, directly delete the elements whose subscripts are in the set T from the S and B sets;

Step 5: Repeat steps 2-4 until set B is an empty set;

Step 6: Return the final result sets D and F.

Preferably, the filtering of the detection frame comprises the following steps:

Step 1: First obtain the network output detection frame coordinate information set B={b ₁ , b ₂ ...b _n }, and the confidence set is S={s ₁ , s ₂ ...s _n }, where b _t , s _t corresponds to the same detection frame coordinate information and confidence;

Step 2: Take out the element with the highest confidence from S, and mark it as i;

Step 3: Get s _i and b _i according to subscript i, and delete s _i and b _i from sets S and B respectively;

和/>

中，对于集合/>

和集合/>

中的每个元素，以检测框的置信度为权重，使用检测框的坐标信息对检测框b_i进行加权，得到/>

然后从集合S中删除集合/>

中元素，从集合B中删除集合/>

中元素；Step 4: traverse the remaining detection frames in the set B, and add the coordinate information and confidence corresponding to all detection frames with an intersection ratio greater than 0.5 with the detection frame b _i to the set respectively

and />

in, for the set />

and collection />

For each element in , the confidence of the detection frame is used as the weight, and the coordinate information of the detection frame is used to weight the detection frame _bi to obtain />

Then delete the set /> from the set S

middle element, remove the set from set B />

middle element;

和s_i加入到结果集D和F中；Step 5: Put

and s _i are added to the result sets D and F;

Step 6: Repeat steps 2-5 until set B is an empty set;

Step 7: Return the final result sets D and F.

Since the present invention uses a compressed feature extraction network, adds an adaptive module to the FocalLoss loss function, and optimizes the filtering of the detection frame, it can simultaneously have higher detection speed and higher accuracy.

Description of drawings

FIG. 1 is a schematic flowchart of a method for optimizing and accelerating a target detection model for a scene with few categories according to the present invention.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Referring to FIG. 1, it shows a schematic flowchart of a method and device for optimizing and accelerating a target detection model for a scene with few categories provided by an embodiment of the present invention. The method includes the following steps:

S100. Acquire and label the picture to be detected.

S110, adding the marked picture to be detected to the streamlined feature extraction device for feature extraction, finally obtaining three sets of feature maps of a specific size as input to the prediction device.

S120, after the features of the picture are extracted by the feature extraction device, the obtained three sets of fixed-size feature maps are sent to the prediction device with an adaptive FocalLoss loss function for result prediction.

S130, after obtaining the prediction result of the network, the result cannot be directly used as an output, because in the result, for the same object to be detected, multiple detection frames are often output, so the main function of the post-processing device is to filter the detection frame, so that for the same object to be detected For the object to be detected, only one detection frame is output.

The device of the embodiment of the present invention has three components: a simplified feature extraction device, a prediction device with an adaptive FocalLoss loss function, and a post-processing device.

Each of the above-mentioned steps and devices is described in detail below:

S100. Acquire and label the picture to be detected.

The picture to be recognized is the picture of the target scene containing a few categories of targets to be detected. First, determine the categories to be detected, such as pedestrians, vehicles, animals, tables, stools, etc.

After the category is determined, the picture is marked according to this. The marking rule is: use the smallest rectangular frame whose sides are parallel to the picture to include all the objects. Generally speaking, each rectangular frame can be used with four free Degree vector representation, one form of expression is (x1, x2, y1, y2), where x1, y1 represent the coordinates of the upper left corner of the rectangular label box, x2, y2 represent the coordinates of the lower right corner of the rectangular label box; another expression The form is (x, y, w, h), where x, y represent the coordinates of the center point of the rectangular label box, and w and h represent the width and height of the rectangular label box, respectively. Obviously these two coordinate representations can be easily converted.

S110, adding the marked picture to be detected to the streamlined feature extraction device for feature extraction, and finally obtaining three groups of feature maps of specific sizes as the input of the prediction device.

The streamlined feature extraction device is responsible for providing a streamlined image feature extraction network with fewer model parameters and faster operation speed.

The feature extraction network is mainly composed of a convolutional layer and a pooling layer. Common feature extraction networks such as: AlexNet, VGG, GoogLeNet, ResNet, SENet, DenseNet, DarkNet, etc. For scenes with high precision requirements, ResNet is often used as the Feature extraction network, and for networks with high speed requirements, DarkNet can be used for feature extraction of pictures, such as YOLO single-stage detector. The application of this device is a real scene, so the network structure of DarkNet53 is used as the basic structure of the feature extraction network of this device.

DaekNet53 has 53 convolutional layers, and these convolutional layers are all composed of 1×1, 3×3 convolution kernels. Each convolution module will be followed by a Batch Normalization layer, which is responsible for speeding up the network during training. Convergence speed; and a LeakyReLU layer as an activation function to increase the nonlinearity of the network. The feature extraction network has a total of 5 downsampling processes. After each downsampling, it consists of 1-8 residual modules. Among them, the residual The module is composed of a convolution module with a convolution kernel size of 3×3 and a convolution module with a convolution kernel size of 1×1, and a residual module is formed after adding skip connections. The purpose of using the residual module is that even if the depth of the network is deepened, it can be easily converged.

Since the ordinary DarkNet network structure will output features with a channel number of 1024 in the last residual module, this is redundant for target detection scenarios that only need to predict fewer categories, so we make the following changes without changing the accuracy On the premise of a slight loss or a slight loss, the calculation amount of the network is reduced, and the operating speed of the feature extraction device is improved.

For all convolution layers, the number of convolution kernels is reduced to X times the original, and the optional value of X is X={1/2,1/4,1/6,1/8}, when X = 1/2, the parameter quantity of the reduced feature extraction sub-network is reduced to 1/2 of the original, and the calculation amount is reduced to 1/4 of the original. It can be seen that as the value of X continues to decrease, the feature extraction The amount of calculations and parameters of the sub-network is continuously decreasing, and the operation speed of the network is continuously accelerating. According to the different requirements of the device for real-time performance, different X parameters can be selected as needed.

It can be clearly seen that in addition to DarkNet, other common feature extraction networks such as: AlexNet, VGG, GoogLeNet, ResNet, SENet, DenseNet, etc., can also be compressed by the above means to speed up network operation.

Adjust the size of the picture to a resolution of N×N, and the selectable value of N is a multiple of 32 between 320-1280. If the size of the input picture is 416×416 pixels, after the picture passes through the feature extraction device, it will pass through 5 downsampling by 2 times, the size of the feature map is: 208×208, 104×104, 52×52, 26×26, 13×13, followed by a feature pyramid structure to predict multiple scale targets, We use the feature maps of the last three layers for multi-scale prediction, that is, connect a small feature conversion network after the feature map sizes of 52×52, 26×26, and 13×13 in the feature extraction network. Except for the difference in the size of the feature maps, the three branch structures connected behind the three feature maps are similar.

S120, after the features of the picture are extracted by the feature extraction device, the obtained three sets of fixed-size feature maps are sent to the prediction device with an adaptive FocalLoss loss function for result prediction.

Prediction device with adaptive FocalLoss loss function: Allows the network to dynamically and automatically adjust hyperparameters, omitting the hyperparameter adjustment process of training the network, so that device users can train high-quality models without mastering deep learning knowledge. model.

The prediction device consists of three parts, namely: a small feature conversion network, a branch for predicting and detecting object classification and coordinates, and a loss function layer.

The small feature conversion network consists of five convolutional layers, and the sizes of the convolution kernels are 1×1, 3×3, 1×1, 3×3, and 1×1. The advantage of this design is that the bottleneck structure has a relatively The amount of calculation is small, and the transformation of the feature space can be fully learned.

The branch of predicting and detecting object classification and coordinates is connected after a small feature transformation network. Optionally, a bottleneck structure composed of 3×3 and 1×1 convolution kernels can be added between these two structures to further deepen the depth of the prediction device. The size of the feature map of this branch is the same as the size of the output feature extraction network, that is, when the output feature map size of the feature extraction network is 13×13, the branch feature map size of the prediction detection target classification and coordinates connected behind it is also 13× 13,

The specific channel number calculation method is as follows:

Determine the total number N of target categories to be identified, generally N is less than or equal to 5, and calculate the channel number C of the output layer feature map according to the formula C=(N+5)×3. The meaning of 5 is: one output represents foreground background prediction, the other four outputs represent coordinate prediction, and 3 represents each output layer, which is responsible for predicting targets of three scales.

For each point in the branch feature map that predicts and detects the target classification and coordinates, it contains preset anchor points of three scales. In the feature map with a size of 13×13, the preset anchor point sizes are 116×90 and 156 respectively. ×198, 373×326, the number of preset anchor points is 13×13×3, which is mainly responsible for the prediction of large-scale targets; in the feature map with a size of 26×26, the preset anchor point sizes are 30×61, 62 ×45, 59×119, the number of preset anchor points is 26×26×3, which is mainly responsible for the prediction of medium-sized targets; in the feature map with a size of 52×52, the preset anchor point sizes are 10×13 and 16 respectively ×30, 33×23, the number of preset anchor points is 52×52×3, which is mainly responsible for the prediction of small-sized targets.

The prediction result is composed of three parts. According to the calculation formula C=(N+5)×3 of the number of branch channels for predicting the detection target classification and coordinates, for each prediction frame, it is composed of N+5 dimensional vectors. First, 1 dimension is determined Whether the target belongs to the foreground; if it is, it means that the prediction frame contains the target to be detected, and the subsequent prediction results are meaningful. The following 4-dimensional vector represents the coordinates of the predicted detection frame, and the maximum value in the subsequent N-dimensional vector represents the target The predicted classification result of ; if not, it means that the predicted frame is the background area of the picture.

The function of the loss function layer is to calculate the distance between the predicted value of the network model and the real value, and optimize the parameters of the network model according to the back propagation algorithm. Therefore, the loss function layer is connected after the branch of predicting and detecting the target classification and coordinates. The input of the loss function layer consists of two parts: the prediction result and the real label value.

When calculating the loss function, a total of three parts of the loss value need to be calculated;

The first step is according to the formula (1), if the predicted target is the foreground area, then calculate the loss of the predicted value and the real value coordinates, using the mean square error loss, the purpose of calculating the root sign at the width w and height h is to reduce The contribution of the width and height prediction results to the entire loss function highlights the importance of the correct prediction of the target center point coordinates x, y.

In the second step, according to the formulas (2) and (3), it is calculated whether there is a target loss, using the adaptive FocalLoss loss function. First, when the actual value of the predicted point is the foreground, that is, y=1, p _t =p . Since the background area in the picture is much larger than the foreground area, the proportion of the foreground and background areas is very unbalanced, so the loss function of this device needs to increase the coefficient α that can adjust the weight ratio of the foreground and background loss on the basis of the traditional cross entropy. Larger, the greater the weight of the foreground area loss; at the same time, in the process of network training, the error value of simple target prediction will become smaller and smaller. In order for the network to learn the prediction of difficult targets, the loss function needs to be increased to increase the difficulty The coefficient γ of the sample loss weight, the larger the value of γ, the greater the weight of the difficult sample loss. The advantage of this design is that it can increase the learning ability of the network for difficult samples, make the prediction results contain more real values, and improve the recall rate of the detection results.

Since the parameter settings of α and γ are very dependent on the device user's understanding of the network, in order to ensure that anyone can easily use the device without any prior knowledge, this device adds an adaptive module, the specific method is as follows:

For each prediction result of each picture, calculate the cross entropy loss, accumulate all the losses that are actually marked as the foreground, denoted as M, and accumulate all the losses that are truly marked as the background, denoted as N;

Calculate log(N/M) and record it as S. If S is less than zero, set S to zero. If it is the first iteration, let the value of S ₀ be S. The purpose of this step is to calculate the ratio of negative sample loss to positive sample loss in the current stage of network training. Generally, more samples in the early stage of network training will be predicted as negative samples, and the value of S will be relatively large;

According to the value of S, calculate the value of α, α=1/(2 ^S +0.2);

The value of γ is calculated according to the values of S and S ₀ , and the calculation formula is γ=min(S ₀ −S,2).

After the above steps, the network can automatically adjust the values of the two hyperparameters α and γ according to the training. Since the network has many prediction errors in the initial training stage, in order to enable the network to converge to a relatively good situation as soon as possible, it is necessary to First, increase the weight of negative samples, and at the same time reduce the weight of learning difficult samples, that is, increase α and reduce γ; therefore, the network will soon reach the first relatively stable stage. At this time, reduce the weight of negative sample punishment and increase positive samples. The penalty of the sample can make the network better learn the really meaningful area, and at this stage S will tend to be stable, so α will also be stable above 0.5, and will not exceed 0.85 at the same time; in order to ensure the stability in the later stage of training , we don't want γ to have a large value, because γ=2 is enough for hard sample learning.

S130, since the detection device is based on a dense anchor point prediction method, after obtaining the prediction result of the network, it cannot directly output the result, because in the result, multiple detection frames are often output for the same target to be detected, so The main function of the post-processing device is to filter the detection frame, so that for the same object to be detected, only one detection frame is output.

Optionally, the following two post-processing methods can be selected according to the different requirements of precision and recall in different scenarios, which are higher precision Precise-NMS (non-maximum suppression algorithm) and higher recall Rate of Better-NMS (Non-Maximum Suppression Algorithm).

The purpose of Precise-NMS is to ensure higher accuracy, while allowing some missed detections, that is, to ensure that the detected results are as correct as possible. Specific steps are as follows:

Step 1: First obtain the network output detection frame coordinate information set B={b ₁ , b ₂ ...b _n }, and the confidence set is S={s ₁ , s ₂ ...s _n }, where b _t , s _t corresponds to the same detection frame coordinate information and confidence;

Step 2: Take out the element with the highest confidence from S, and mark it as i;

Step 3: Get s _i and b _i according to subscript i, and delete s _i and b _i from sets S and B respectively;

Step 4: Traverse the remaining detection frames in the set B, and add the subscripts of all detection frames with an intersection ratio greater than 0.5 with _bi to the set T. If T is not an empty set, add bi _and s _i to In the result sets D and F, delete the elements whose subscripts are in the set T from the S and B sets at the same time, if T is an empty set, directly delete the elements whose subscripts are in the set T from the S and B sets;

Step 5: Repeat steps 2-4 until set B is an empty set;

Step 6: Return the final result sets D and F.

The purpose of this strategy is that if there are multiple anchor points biased towards an object for learning at the same time, and at the same time they can give the object a relatively high score in the detection stage, it can be considered that there is a greater possibility of an object in this area.

The meaning of the intersection ratio mentioned above is: the ratio of the area of the intersection of the two detection frames and the area of the union.

Better-NMS can get a higher quality detection frame, making the detection result have a higher recall rate. Specific steps are as follows:

Step 1: First obtain the network output detection frame coordinate information set B={b ₁ , b ₂ ...b _n }, and the confidence set is S={s ₁ , s ₂ ...s _n }, where b _t , s _t corresponds to the same detection frame coordinate information and confidence;

Step 2: Take out the element with the highest confidence from S, and mark it as i;

Step 3: Get s _i and b _i according to subscript i, and delete s _i and b _i from sets S and B respectively;

和/>

中，对于集合/>

和集合/>

中的每个元素，以检测框的置信度为权重，使用检测框的坐标信息对检测框b_i进行加权，得到/>

然后从集合S中删除集合/>

中元素，从集合B中删除集合/>

中元素；Step 4: traverse the remaining detection frames in the set B, and add the coordinate information and confidence corresponding to all detection frames with an intersection ratio greater than 0.5 with the detection frame b _i to the set respectively

and />

in, for the set />

and collection />

For each element in , the confidence of the detection frame is used as the weight, and the coordinate information of the detection frame is used to weight the detection frame _bi to obtain />

Then delete the set /> from the set S

middle element, remove the set from set B />

middle element;

和s_i加入到结果集D和F中；Step 5: Put

and s _i are added to the result sets D and F;

Step 6: Repeat steps 2-5 until set B is an empty set;

Step 7: Return the final result sets D and F.

The weighting mentioned above is calculated according to the following formula:

The purpose of this strategy is to reconsider the role of those high-scoring detection frames that are filtered out, because those filtered out detection frames are also relatively good results obtained by the network after training and learning, which also includes objects. Semantic information, so that the weighted result can get a higher quality detection frame.

Claims (3)

1. A method for optimizing and accelerating a target detection model for a few-category scene, comprising the following steps: Obtain and label the image to be detected; The marked pictures are added to the feature extraction device for feature extraction, and three sets of fixed-size feature maps are obtained; Send three sets of fixed-size feature maps to the prediction device with FocalLoss loss function for result prediction; Filter the detection frame of the prediction result so that only one detection frame is output for the same object to be detected; It is characterized in that an adaptive module is added to the FocalLoss loss function, specifically: For each prediction result of each picture, calculate the cross entropy loss, accumulate all the losses that are actually marked as the foreground, denoted as M, and accumulate all the losses that are truly marked as the background, denoted as N; Calculate log(N/M), record it as S, if S is less than zero, then S is set to zero, if it is the first iteration, let the value of S ₀ be S; According to the value of S, calculate the value of α, α=1/(2 ^S +0.2); The value of γ is calculated according to the values of S and S ₀ , and the calculation formula is γ=min(S ₀ −S,2). 2. the method for the object detection model optimization and acceleration of few category scenes as claimed in claim 1, it is characterized in that the filtering of described detection frame comprises the following steps: Step 1: First obtain the network output detection frame coordinate information set B={b ₁ , b ₂ ...b _n }, and the confidence set is S={s ₁ , s ₂ ...s _n }, where b _t , s _t corresponds to the same detection frame coordinate information and confidence; Step 2: Take out the element with the highest confidence from S, and mark it as i; Step 3: Get s _i and b _i according to subscript i, and delete s _i and b _i from sets S and B respectively; Step 4: Traverse the remaining detection frames in the set B, and add the subscripts of all detection frames whose intersection ratio with _bi is greater than 0.5 to the set _{T. If T is not an empty set, add bi and s i respectively} To the result sets D and F, delete the elements whose subscripts are in the set T from the S and B sets at the same time; Step 5: Repeat steps 2-4 until set B is an empty set; Step 6: Return the final result sets D and F. 3. the method for optimizing and accelerating the target detection model for few category scenes as claimed in claim 1, it is characterized in that the filtering of described detection frame comprises the following steps: Step 1: First obtain the network output detection frame coordinate information set B={b ₁ , b ₂ ...b _n }, and the confidence set is S={s ₁ , s ₂ ...s _n }, where b _t , s _t corresponds to the same detection frame coordinate information and confidence; Step 2: Take out the element with the highest confidence from S, and mark it as i; Step 3: Get s _i and b _i according to subscript i, and delete s _i and b _i from sets S and B respectively; Step 4: traverse the remaining detection frames in the set B, and add the coordinate information and confidence corresponding to all detection frames with an intersection ratio greater than 0.5 with the detection frame b _i to the set respectively

and />

in, for the set />

and collection />

For each element in , the confidence of the detection frame is used as the weight, and the coordinate information of the detection frame is used to weight the detection frame _bi to obtain />

Then delete the set /> from the set S

middle element, remove the set from set B />

middle element; Step 5: Put

and s _i are added to the result sets D and F; Step 6: Repeat steps 2-5 until set B is an empty set; Step 7: Return the final result sets D and F.

Images