CN111950551A - A target detection method based on convolutional neural network - Google Patents

Abstract

The invention relates to a target detection method based on a convolutional neural network, comprising: performing feature extraction based on a residual convolutional neural network to obtain a layer-by-layer basic feature map; and sequentially fusing the basic feature maps from shallow to deep to obtain fusion features Figure; perform candidate frame extraction on the fusion feature map based on the region generation network to obtain a candidate target region feature map; obtain a region of interest feature map according to the fusion feature map and the candidate target region feature map; The region feature maps are based on fully convolutional layers to obtain classification scores and bounding box regression. The present invention has higher detection accuracy for small targets and occluded targets.

Description

A target detection method based on convolutional neural network

technical field

The invention relates to the technical field of image information processing, in particular to a target detection method based on a convolutional neural network.

Background technique

With the increasing pressure of road traffic, the intelligent management and control of road vehicles through computer technology has become a research hotspot; the use of road monitoring equipment to detect vehicle targets, and mastering the vehicle data and driving trajectories of the road network is the key to optimizing traffic and alleviating traffic. The premise of pressure, and vehicle target detection is the research basis in the fields of unmanned driving, vehicle tracking, and vehicle feature recognition.

At present, convolutional neural networks are widely used in the field of vehicle target detection. The commonly used ones are generally divided into single-stage detection algorithms and two-stage detection algorithms. The single-stage detection algorithm is a regression-based target detection algorithm. The two-stage detection algorithm first generates The candidate regions are then classified and refined. Due to the difference in algorithm structure, the two-stage detection algorithm has higher detection accuracy, but the detection speed is lower than that of the single-stage detection algorithm, which is suitable for scenarios that require higher detection accuracy.

The existing two-stage target detection algorithms have the following problems: due to the few features of occluded targets and small targets, the existing algorithms do not fully utilize the shallow position information and context information, which makes the detection accuracy of small targets and occluded targets relatively low. Low.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a target detection method based on a convolutional neural network with high detection accuracy for small targets and occluded targets.

For achieving the above object, the present invention provides the following scheme:

A target detection method based on convolutional neural network, including:

Feature extraction based on residual volume neural network to obtain basic feature map layer by layer;

The basic feature maps are fused sequentially from shallow to deep to obtain a fusion feature map;

Extracting candidate frames from the fusion feature map based on the region generation network to obtain a candidate target region feature map;

Obtain a region of interest feature map according to the fusion feature map and the candidate target region feature map;

The classification score and bounding box regression are obtained based on the fully convolutional layer according to the region of interest feature map.

Preferably, the basic feature map includes a first feature map, a second feature map, a third feature map and a fourth feature map.

Preferably, the basic feature maps are fused sequentially from shallow to deep to obtain a fusion feature map, including:

Perform downsampling processing on the first feature map to obtain a downsampling feature map;

Performing convolution dimension reduction processing on the second feature map to obtain a dimension reduction feature map, where the number of channels of the dimension reduction feature map is the same as the number of channels of the downsampling feature map;

The down-sampling feature map and the dimension reduction feature map are fused to obtain an initial fused feature map; similarly, the fused feature map is finally obtained.

Preferably, performing down-sampling processing on the first feature map to obtain a down-sampling feature map, including:

Perform down-sampling processing on the first feature map based on n branch hole convolutions; n is a positive integer greater than 1;

The down-sampling feature map is obtained by fusing the first feature map subjected to down-sampling processing through hole convolution of each branch.

Preferably, the n is 3, and the void ratios of the three branches are 1, 2, and 3, respectively.

Preferably, the region-based generation network performs candidate frame extraction on the fusion feature map to obtain a candidate target region feature map, including:

Perform convolution processing on the fusion feature map based on the first set convolution kernel to obtain a first convolution feature map;

Perform convolution processing on the first convolution feature map based on the second set convolution kernel to obtain a second convolution feature map;

Perform convolution processing on the second convolution feature map based on the second set convolution kernel to obtain a third convolution feature map;

The second convolutional feature map and the third convolutional feature map are respectively input into two parallel fully-connected layers, and processed based on the set anchor frame to obtain the candidate target region feature map.

Preferably, the classification score and the bounding box regression are obtained based on the fully convolutional layer according to the feature map of the region of interest, including:

Obtaining the initial classification score and initial frame regression based on the fully convolutional layer according to the region of interest feature map;

Replace the set anchor frame with the initial frame regression, and perform subsequent steps in turn, by setting m thresholds, and repeating this process m times to obtain the classification score and the frame regression; m is greater than or A positive integer equal to 1.

Preferably, the first set convolution kernel is 3×3; the second set convolution kernel is 1×1.

Preferably, obtaining the region of interest feature map according to the fusion feature map and the candidate target region feature map includes:

Based on ROIAlign, the fusion feature map and the candidate target region feature map are fused to obtain the initial region of interest feature map;

Enlarging the initial region-of-interest feature map according to a set multiple to obtain an enlarged region-of-interest feature map;

Perform global context extraction on the initial region of interest feature map based on the enlarged region of interest feature map to obtain context information;

The initial region of interest feature map is fused with the context information based on ROIAlign to obtain the region of interest feature map.

Preferably, the residual volume neural network is a ResNet-101 network.

According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

The invention relates to a target detection method based on a convolutional neural network, comprising: performing feature extraction based on a residual convolutional neural network to obtain a layer-by-layer basic feature map; and sequentially fusing the basic feature maps from shallow to deep to obtain fusion features Figure; perform candidate frame extraction on the fusion feature map based on the region generation network to obtain a candidate target region feature map; obtain a region of interest feature map according to the fusion feature map and the candidate target region feature map; The region feature maps are based on fully convolutional layers to obtain classification scores and bounding box regression. The present invention has higher detection accuracy for small targets and occluded targets.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings required in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some of the present invention. In the embodiments, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative labor.

FIG. 1 is a flowchart of a target detection method based on a convolutional neural network according to the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

The purpose of the present invention is to provide a target detection method based on a convolutional neural network with high detection accuracy for small targets and occluded targets.

In order to make the above objects, features and advantages of the present invention more clearly understood, the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

Fig. 1 is the flow chart of the target detection method based on the convolutional neural network of the present invention. As shown in Fig. 1, the present invention provides a target detection method based on the convolutional neural network, including:

In step S1, feature extraction is performed based on the residual volume neural network ResNet-101 to obtain a layer-by-layer basic feature map, which specifically includes a first feature map, a second feature map, a third feature map, and a fourth feature map. In this embodiment, the details of each convolutional layer of the ResNet-101 are shown in Table 1.

Table 1. Each convolutional layer of ResNet-101

where w is the width of the region of interest and h is the height of the region of interest.

Step S2, the basic feature maps are fused sequentially from shallow to deep to obtain a fused feature map.

Taking the fusion of the first feature map and the second feature map as an example to illustrate, the specific process is as follows:

The first feature map is down-sampled based on n branch hole convolutions; n is a positive integer greater than 1. In this embodiment, n is set to 3, the size of the convolution kernel is 3×3, and the convolution step size is 2;

The down-sampling feature map is obtained by fusing the first feature map subjected to down-sampling processing through hole convolution of each branch. The specific calculation formula is:

F=H _3,1(x) +H _3,2(x) +H _3,3(x) ;

In the formula: F represents the fused down-sampling feature map, H _{k, r, (x)} represents the hole convolution, k represents the size of the convolution kernel, r represents the hole rate, and x is the first feature map.

A 1×1 convolution kernel is used to perform convolution dimension reduction processing on the second feature map to obtain a dimension-reduced feature map, where the number of channels of the dimension-reduced feature map is the same as the number of channels of the down-sampling feature map.

The down-sampled feature map and the dimension-reduced feature map are fused to obtain an initial fusion feature map.

The fusion feature map is obtained by performing fusion in sequence according to the above steps.

Step S3, extracting candidate frames from the fusion feature map based on the region generation network to obtain a feature map of candidate target regions.

As an optional implementation manner, the step S3 of the present invention includes:

Step S31 , performing convolution processing on the fusion feature map based on the first set convolution kernel to obtain a first convolution feature map. In this embodiment, the size of the first set convolution kernel is 3×3.

Step S32, performing convolution processing on the first convolution feature map based on the second set convolution kernel to obtain a second convolution feature map. In this embodiment, the size of the second set convolution kernel is 1×1.

Step S33, performing convolution processing on the second convolution feature map based on the second set convolution kernel to obtain a third convolution feature map.

Step S34, the second convolution feature map and the third convolution feature map are respectively input into two parallel fully connected layers, and processed based on the set anchor frame to obtain the candidate target region feature map.

Step S4, obtaining a region of interest feature map according to the fusion feature map and the candidate target region feature map.

Specifically, the step S4 includes:

Step S41 , fuse the fusion feature map and the candidate target region feature map based on the ROI Align to obtain an initial region of interest feature map.

Step S42: Enlarging the initial region-of-interest feature map according to a set multiple to obtain an enlarged region-of-interest feature map. In this embodiment, the set multiple is 1.5.

Step S43 , perform global context extraction in four directions of up, down, left, and right on the initial region of interest feature map based on the enlarged region of interest feature map to obtain context information.

Step S44, based on ROIAlign, map the initial region of interest feature map and the context information into a rectangular frame of the same size, and perform fusion to obtain the region of interest feature map.

Step S5, obtaining classification score and bounding box regression based on the fully convolutional layer according to the feature map of the region of interest.

Specifically, the initial classification score and the initial bounding box regression are obtained based on the fully convolutional layer according to the feature map of the region of interest.

Replace the set anchor frame with the initial frame regression, and perform the following steps in turn, by setting m thresholds and repeating this process m times to obtain the classification score and the frame regression; m is greater than or a positive integer equal to 1. In this embodiment, m is taken as 3, and the three thresholds are respectively 0.5, 0.6 and 0.7.

The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the various embodiments can be referred to each other.

In this paper, specific examples are used to illustrate the principles and implementations of the present invention. The descriptions of the above embodiments are only used to help understand the methods and core ideas of the present invention; meanwhile, for those skilled in the art, according to the present invention There will be changes in the specific implementation and application scope. In conclusion, the contents of this specification should not be construed as limiting the present invention.

Claims (10)

1. a target detection method based on convolutional neural network, is characterized in that, comprises: Feature extraction based on residual volume neural network to obtain basic feature map layer by layer; The basic feature maps are fused sequentially from shallow to deep to obtain a fusion feature map; Extracting candidate frames from the fusion feature map based on the region generation network to obtain a candidate target region feature map; Obtain a region of interest feature map according to the fusion feature map and the candidate target region feature map; The classification score and bounding box regression are obtained based on the fully convolutional layer according to the region of interest feature map. 2 . The target detection method based on a convolutional neural network according to claim 1 , wherein the basic feature map comprises a first feature map, a second feature map, a third feature map and a fourth feature map. 3 . . 3. a kind of target detection method based on convolutional neural network according to claim 2, is characterized in that, described basic feature map is fused successively from shallow to deep, obtains fusion feature map, comprising: Perform downsampling processing on the first feature map to obtain a downsampling feature map; Performing convolution dimension reduction processing on the second feature map to obtain a dimension reduction feature map, where the number of channels of the dimension reduction feature map is the same as the number of channels of the downsampling feature map; The down-sampling feature map and the dimension reduction feature map are fused to obtain an initial fused feature map; similarly, the fused feature map is finally obtained. 4. The target detection method based on a convolutional neural network according to claim 3, wherein the downsampling process is performed on the first feature map to obtain a downsampling feature map, comprising: Perform down-sampling processing on the first feature map based on n branch hole convolutions; n is a positive integer greater than 1; The down-sampling feature map is obtained by fusing the first feature map subjected to down-sampling processing through hole convolution of each branch. 5 . The target detection method based on a convolutional neural network according to claim 4 , wherein the n is 3, and the void ratios of the three branches are 1, 2, and 3, respectively. 6 . 6. A target detection method based on a convolutional neural network according to claim 1, wherein the region-based generation network performs candidate frame extraction on the fusion feature map to obtain a candidate target region feature map, comprising: : Perform convolution processing on the fusion feature map based on the first set convolution kernel to obtain a first convolution feature map; Perform convolution processing on the first convolution feature map based on the second set convolution kernel to obtain a second convolution feature map; Perform convolution processing on the second convolution feature map based on the second set convolution kernel to obtain a third convolution feature map; The second convolutional feature map and the third convolutional feature map are respectively input into two parallel fully-connected layers, and processed based on the set anchor frame to obtain the candidate target region feature map. 7. a kind of target detection method based on convolutional neural network according to claim 6, is characterized in that, described according to described region of interest feature map to obtain classification score and bounding box regression based on full convolution layer, comprising: Obtaining the initial classification score and initial frame regression based on the fully convolutional layer according to the region of interest feature map; Replace the set anchor frame with the initial frame regression, and perform subsequent steps in sequence, by setting m thresholds, and repeating this process m times to obtain the classification score and the frame regression; m is greater than or A positive integer equal to 1. 8 . The target detection method based on a convolutional neural network according to claim 6 , wherein the first set convolution kernel is 3×3; the second set convolution kernel is 1 8 . ×1. 9. A target detection method based on a convolutional neural network according to claim 1, wherein the obtaining a region of interest feature map according to the fusion feature map and the candidate target region feature map, comprising: Based on ROI Align, the fusion feature map and the candidate target region feature map are fused to obtain an initial region of interest feature map; Enlarging the initial region-of-interest feature map according to a set multiple to obtain an enlarged region-of-interest feature map; Perform global context extraction on the initial region of interest feature map based on the enlarged region of interest feature map to obtain context information; Based on the ROI Align, the initial region of interest feature map and the context information are fused to obtain the region of interest feature map. 10 . The target detection method based on a convolutional neural network according to claim 1 , wherein the residual convolutional neural network is a ResNet-101 network. 11 .

Images