Target detection method based on small target contrast feature enhancement

Abstract

The invention discloses a target detection method based on small target contrast feature enhancement ST-CFE. The method comprises the following steps: establishing a small target contrast feature enhancement ST-CFE network; inputting the pictures into a backbone network to obtain feature maps of different scales; the feature map is sent to a fast spatial pyramid pooling cross-layer fusion module SPPF-CLFM; the feature map is sent to a bidirectional feature pyramid network BiFPN to promote multi-scale fusion; the feature map in the BiFPN is sent to a small target comparison feature enhancement module ST-CFEM to weaken large target information in the shallow feature map, and small target feature information expression is compared and enhanced; the enhanced feature map is sent to an efficient channel attention decoupling detection head (ECA-DDH); performing non-maximum suppression (NMS) operation on the detail information and the position information of the target to obtain a final output result; the method has high performance in the aspects of accuracy, small target detection comprehensiveness and small target positioning.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and target detection, and in particular to a target detection method based on small target contrast feature enhancement. Background Art

[0002] Due to the special environment of the elevated warehouse area, the smallest safety helmet in the acquired monitoring image accounts for only 0.004% of the pixels, and the larger safety helmet accounts for less than 0.8% of the pixels, which are extremely small targets. In the field of target detection, detecting small targets has always been a very challenging task, because small targets often occupy a small area in the image and have complex and varied forms of expression, which increases the difficulty of detection. In the development of target detection based on deep learning, research on small target detection has made some achievements. At the beginning, the method to improve the performance of small target detection was mainly super-resolution. Zhang et al. [1] The generator is used to generate super-resolution images, and a multi-task discriminator network is introduced to guide the generator network to produce super-resolution images.

[0003] Feature Pyramid Network (FPN) [2] An effective multi-scale framework is constructed to solve the problem of low performance of small object detection. In FPN, multi-scale feature maps are created by propagating semantic information from high-order to low-order. Multi-scale feature description is achieved by fusing shallow features and deep features, which greatly improves the performance of object detection. Extended Feature Pyramid Network (E-FPN) [3] Using additional detection layers, a high-resolution layer is tailored for detecting small objects to improve detection accuracy. Zhang et al. [4] Add feature selection and feature alignment modules to the lateral connection part of the feature pyramid to enhance the important features of pedestrians in the input image and enhance the model's multi-scale feature fusion capability for small target pedestrians. Chen et al. [5] By adding a lateral connection module based on pixel shuffle to expand the number of channels and aggregate channel information, the feature information in the channel is retained while reducing the channel dimension, alleviating the information loss problem caused by the reduction of the number of channels during the upsampling process of FPN.

[0004] Although the above methods fully utilize the characteristics of shallow features that are rich in more detailed information, they ignore the coexistence of large and small target information in shallow features. The detector will confuse the semantic information of small and large targets, resulting in poor network detection performance for small targets. Moreover, most networks only supplement the detailed information under different receptive fields in the channel dimension, ignoring the impact of detailed information in the spatial dimension on the positioning accuracy of small targets, resulting in low network accuracy. Since the classification and regression tasks share the same feature information, the coupled detection head cannot fully utilize the feature differences between different tasks.

[0005] In summary, the traditional method of detecting small targets has the following defects:

[0006] (1) The information loss in the spatial dimension during the feature extraction process will affect the positioning accuracy of small targets;

[0007] (2) When using shallow feature layers to detect small targets, due to the coexistence of large and small target information in the shallow feature layer, the detector will confuse the semantic information of small targets and large targets, resulting in poor performance of the network for small target detection. At this time, the large target information in the shallow feature layer can be regarded as noise that interferes with the extraction of small target features;

[0008] (3) In the final feature extraction process of two-stage object detection, there is a problem of mutual interference between classification and regression tasks. That is, the two tasks focus on different areas of interest. Classification pays more attention to which category the extracted features are closest to, while regression pays more attention to the position coordinates of the predicted box and the true annotation box.

[0009] [1]Bai Y, Zhang Y, Ding M, et al. Sod-mtgan: Small object detection via multi-task generative adversarial network [C] / / Proceedings of the European Conference on Computer Vision (ECCV). 2018: 206-221.

[0010] [2]Lin TY, Dollár P, Girshick R, et al.Feature pyramid networks for object detection[C] / / Proceedings of the IEEE conference on computer vision and pattern recognition.2017:2117-2125.

[0011] [3]Deng C,Wang M,Liu L,et al.Extended feature pyramid network for small object detection[J].IEEE Transactions on Multimedia,2021,24:1968-1979.

[0012] [4]Zhang Y, Zhang S, Xin D, et al.A Small Target Pedestrian DetectionModel Based on Autonomous Driving[J]. Journal of Advanced Transportation, 2023.

[0013] [5]Chen S, Zhao J, Zhou Y, et al.Info-FPN: An informative feature pyramidnetwork for object detection in remote sensing images[J].Expert Systems withApplications,2023,214:119-132. Summary of the invention

[0014] In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a target detection method based on small target contrast feature enhancement, which has high performance in terms of accuracy, comprehensiveness of small target detection and small target positioning.

[0015] To achieve the above object, the technical solution adopted by the present invention is: a target detection method based on small target contrast feature enhancement, comprising the following steps:

[0016] Step 1: Build a ST-CFE network based on small target contrast feature enhancement;

[0017] Step 2: Input the image into the backbone network CSPDarknet53 of the small target contrast feature enhancement ST-CFE network to obtain feature maps of different scales;

[0018] Step 3: Input the last feature map of the backbone network into the fast spatial pyramid pooling cross-layer fusion module SPPF-CLFM, and use the fast spatial pyramid pooling cross-layer fusion module SPPF-CLFM to capture the target information at multiple scales and reduce the target information loss in the spatial dimension;

[0019] Step 4: Add a small target detection head to the shallow feature map with high resolution and more position and detail information, input the shallow feature map of the backbone network and the output feature map of the fast spatial pyramid pooling cross-layer fusion module SPPF-CLFM into the bidirectional feature pyramid network BiFPN, and use the bidirectional feature pyramid network BiFPN to promote multi-scale feature fusion, avoid the loss of feature information, and retain the detail information of small target features;

[0020] Step 5: Input the feature map in the bidirectional feature pyramid network BiFPN into the small target contrast feature enhancement module ST-CFEM. Through the small target contrast feature enhancement module ST-CFEM, the large target information in the shallow feature map is weakened, and the small target feature expression is contrast enhanced;

[0021] Step 6: Build an efficient channel attention decoupling detection head ECA-DDH, and input the output of the small target contrast feature enhancement module ST-CFEM and the last two layers of feature maps of the bidirectional feature pyramid network BiFPN into the efficient channel attention decoupling detection head ECA-DDH;

[0022] Step 7: Perform non-maximum suppression (NMS) on the target category and location information output by the efficient channel attention decoupling detection head ECA-DDH to obtain the final output result.

[0023] As a further improvement of the present invention, in step 3, the fast spatial pyramid pooling cross-layer fusion module SPPF-CLFM uses continuous maximum pooling and channel splicing to capture target information at different scales, and uses skip connections and pixel-by-pixel addition to reduce the loss of target information in the spatial dimension.

[0024] As a further improvement of the present invention, the step 5 is specifically as follows:

[0025] For the shallow feature map, two-way parallel dilated convolution is used to obtain different receptive fields, and channel attention is used to learn the importance of different channels. Finally, the deep feature map is subtracted from the shallow feature map after feature enhancement to weaken the large target information of the shallow feature map and enhance the expression of small target feature information.

[0026] As a further improvement of the present invention, in step 6, by decoupling the coupled detection head into classification and regression branches, the classification and regression branches are allowed to use different loss functions and weights to better balance the training process of the two branches; the classification and regression branches are allowed to focus on the target semantic information and position information respectively, reducing mutual interference between branches and improving the network's ability to classify and locate targets; and efficient channel attention ECA is added to the classification branch to adaptively learn the importance of each channel without losing feature information, thereby improving the network classification accuracy.

[0027] The beneficial effects of the present invention are:

[0028] The present invention has high performance in terms of accuracy, comprehensiveness of small target detection and small target positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of the structure of a small target contrast feature enhancement network (ST-CFE) in Example 1 of the present invention;

[0030] Figure 2 Schematic diagram of the structure of a fast spatial pyramid pooling cross-layer fusion module (SPPF-CLFM) in Example 1 of the present invention;

[0031] Figure 3 Schematic diagram of the structure of a small target contrast feature enhancement module (ST-CFEM) in Example 1 of the present invention;

[0032] Figure 4 Schematic diagram of the structure of the efficient channel attention decoupling detection head (ECA-DDH) in Example 1 of the present invention. DETAILED DESCRIPTION

[0033] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0034] Definitions of Key Terms:

[0035] 1. Feature Pyramid Network (FPN);

[0036] 2.Spatial Pyramid Pooling Fast (SPPF);

[0037] 3.Spatial Pyramid Pooling Fast-CrossLayer Fusion Module (SPPF-CLFM);

[0038] 4. Small Target Contrast Feature Enhancement Module (ST-CFEM);

[0039] 5. Efficient Channel Attention Decoupled Detection Head (ECA-DDH);

[0040] 6. Bi-directional Feature Pyramid Network (BiFPN);

[0041] 7. Efficient Channel Attention (ECA);

[0042] 8.CSPDarknet53:Cross Stage Partial Darknet 53;

[0043] 9. Non-Maximum Suppression (NMS).

[0044] Example 1

[0045] A target detection method based on small target contrast feature enhancement includes the following steps:

[0046] Step 1: Build a ST-CFE network based on small target contrast feature enhancement.

[0047] like Figure 1As shown in the figure, the small target contrast feature enhancement network consists of a backbone network, a neck network, and a detection head. The image is input into the backbone network CSPDarknet53 to obtain feature maps of different scales (C1, C2, C3, C4, C5), and then the C5 feature map is sent to the fast spatial pyramid pooling cross-layer fusion module (SPPF-CLFM). Three consecutive 5×5 maximum poolings are used to expand the receptive field, and then the output of each maximum pooling is channel-joined with the input to achieve the function of capturing small target information at different scales. Skip connections and pixel-by-pixel addition operations are used to reduce the loss of target information in the spatial dimension, and feature map P5 is obtained. A small target detection head is added to the shallow feature map C2 with high resolution and more position and detail information. The bidirectional feature pyramid network BiFPN is used in the neck network. The bottom-up, top-down and lateral connections are used to transmit feature information to promote the interaction of cross-layer features and improve the network's receptive field and semantic expression ability. However, the large target information in the shallow feature map (C2-C3) will interfere with the extraction of small target information. Therefore, D2, P4 and D3, P5 in BiFPN are sent to the small target contrast feature enhancement module (ST-CFEM). The shallow feature maps D2 and D3 are convolved with two-way parallel convolutions of different void rates to obtain different receptive fields. Then, channel attention is used to obtain more accurate feature information of the feature map in the channel dimension. The feature maps P4 and P5 are upsampled. The shallow feature map is subtracted from the deep feature map to weaken the large target information in the shallow feature map, and the small target feature information is enhanced by contrast, thereby improving the network's detection performance for small targets. After that, the outputs of ST-CFEM and BiFPN are sent to the efficient channel attention decoupled detection head (ECA-DDH) of different scales in turn. The original coupled detection head is decoupled into classification and regression branches, so that the network can learn the target's detail information and location information without interfering with each other. The classification branch adds efficient channel attention (ECA) so that the network can adaptively learn the importance of each channel without losing feature information. Finally, the output of ECA-DDH is subjected to non-maximum suppression (NMS) to obtain the final output result of the network;

[0048] Step 2: Input the image into the backbone network CSPDarknet53 to obtain feature maps of different scales (C1, C2, C3, C4, C5);

[0049] Step 3: Use the fast spatial pyramid pooling cross-layer fusion module (SPPF-CLFM) to capture target information at multiple scales and reduce target information loss in the spatial dimension.

[0050] like Figure 2As shown in Figure 1, the output feature map C5 of the backbone network is sent to the fast spatial pyramid pooling cross-layer fusion module (SPPF-CLFM) to obtain the P5 feature map with multi-scale information. First, a 1×1 convolution is used on the input feature map C5 to realize the information interaction of the input feature map channel dimension and realize feature fusion. A 3×3 convolution is used to extract the local feature information of the feature map to obtain the feature map S2, as shown in formula (1).

[0051] S2=C 3×3 (C 1×1 (C5)) (1)

[0052] Among them C 1×1 and C 3×3 They represent 1×1 convolution and 3×3 convolution respectively.

[0053] The feature map S2 is subjected to a fast spatial pyramid pooling (SPPF) operation, that is, a 1×1 convolution is applied to the feature map S2, and then three consecutive 5×5 maximum poolings are used to capture small target information at different scales, and channel concatenation is used to integrate the target information at these different scales in the channel dimension, and a 1×1 convolution is used to obtain the feature map S3, as shown in formula (2).

[0054] S3=Cat(C 1×1 (MP 5×5 (MP 5×5 (MP 5×5 (C 1×1 (S2)))),C 1×1 (S2))) (2)

[0055] Among them, Cat indicates channel splicing, MP 5×5 represents the maximum pooling of size 5×5, C 1×1 It represents a 1×1 convolution.

[0056] Feature map S3 is convolved with 3×3 to obtain feature map S4, and feature map C5 is convolved with 1×1 to obtain feature map S z The feature map P5 is obtained by adding pixel by pixel to the feature map S4 to complete the information fusion in the spatial dimension, reducing the information loss in the spatial dimension caused by the maximum pooling, so that the network can effectively perceive and locate small targets, as shown in formula (3).

[0057] P5=Add(C 3×3 (S3),C 1×1 (C5)) (3)

[0058] Where Add means adding the feature map pixel by pixel, C 1×1 and C 3×3 They represent 1×1 convolution and 3×3 convolution respectively.

[0059] Step 4: Add a small target detection head to the shallow feature map with high resolution and more position and detail information, and use the bidirectional feature pyramid network BiFPN to promote multi-scale feature fusion, avoid feature information loss, and retain the detail information of small target features. Figure 1 As shown in the neck network in the figure, the feature maps C2, C3, C4 from the backbone network and the output feature map P5 of SPPF-CLFM are sent to the bidirectional feature pyramid network (BiFPN), and the feature map P5 is upsampled and concatenated with the feature map C4 to obtain the feature map P4, the feature map P4 is upsampled and concatenated with the feature map C3 to obtain the feature map P3, and the feature map C2 is concatenated with the feature map P3 to obtain the feature map D2, completing the bottom-up process of BiFPN. After that, the feature map D2 is downsampled and concatenated with the feature map P3 and the feature map C3 to obtain the feature map D3, the feature map D3 is downsampled and concatenated with the feature map P4 and the feature map C4 to obtain the feature map D4, and finally the feature map D4 is downsampled and concatenated with the feature map P5 to obtain the feature map D5, completing the top-down and lateral jump connection of BiFPN, realizing the multi-scale fusion of features and the utilization of context information.

[0060] Step 5: Through the small target contrast feature enhancement module (ST-CFEM), the large target information in the shallow feature map (D2, D3) is weakened, and the small target feature expression is enhanced. Specifically, two-way parallel dilated convolution is used to obtain different receptive fields for the shallow feature map, and the importance of different channels is learned by channel attention. Finally, the deep feature map is subtracted from the shallow feature map after feature enhancement to weaken the large target information in the shallow feature map and enhance the small target feature information expression.

[0061] like Figure 3 As shown, this module has two inputs: one is the shallow feature map (i is the number of feature layers, i is 2 and 3), the other is the deep feature map First, two convolution kernels with different void rates are used to compare Perform convolution operation to obtain two feature maps with different receptive field sizes, and then perform channel splicing to obtain feature map D' i , the specific operation is shown in formula (4).

[0062]

[0063] Among them, Cat is channel splicing, DC 3,2 、DC 3,3 They represent dilated convolutions with a convolution kernel of 3×3 and dilation rates of 2 and 3 respectively.

[0064] Then the feature map D'i Perform channel attention operation to obtain D d , while the original feature map Through 1×1 convolution and Sigmoid, the channel dimension is increased, and the feature map D" i With D d Multiply pixel by pixel to get feature D' d , the specific operations are shown in formulas (5) and (6).

[0065]

[0066]

[0067] In the formula represents pixel-by-pixel multiplication, CR represents 1×1 convolution plus ReLu activation function, and GAP represents global average pooling.

[0068] In order to weaken the large target information in the shallow feature map and enhance the expression of small target feature information, the feature map Four times upsampling and using 1×1 convolution to integrate features across channels to obtain the feature map P' i+2 . Then P' i+2 and D' d Multiplication, because after the Sigmoid operation, the value of the feature map is between 0 and 1, feature map multiplication can scale the large target interference information to prevent pixel-by-pixel subtraction from affecting subsequent large target detection. The specific operations are shown in equations (7) and (8).

[0069]

[0070]

[0071] Where CS represents 1×1 convolution plus Sigmoid activation function, Up 4 represents four times upsampling, Represents pixel-by-pixel subtraction.

[0072] Step 6. Build an efficient channel attention decoupled detection head (ECA-DDH). By decoupling the coupled detection head into classification and regression branches, the classification and regression branches are allowed to use different loss functions and weights to better balance the training process of the two branches; the classification and regression branches are allowed to focus on the target semantic information and location information respectively, reducing mutual interference between branches and improving the network's ability to classify and locate targets. Add efficient channel attention ECA to the classification branch to adaptively learn the importance of each channel without losing feature information, thereby improving the network classification accuracy.

[0073] like Figure 4 As shown, in this embodiment, the output P of ST-CFEMout The feature maps D4 and D5 of the neck network are sent to ECA-DDH. First, a 1×1 convolution is performed on the input features, and the feature maps with target semantic information and position information are sent to the classification branch and the regression branch. In the classification branch, the input features are sent to the CBL layer (including 3×3 convolution, batch normalization layer and ReLu activation function) to obtain the target semantic information and realize feature normalization. Then, an efficient channel attention ECA operation is performed, that is, global average pooling is performed on the input features, and a 1D convolution with a convolution kernel of 1×5 is used to integrate channel information without reducing the number of channels. Then, a learnable channel weight W is obtained through a Sigmoid activation function, and each channel of the input feature is weighted to adaptively learn the importance of each channel. Finally, a 1×1 convolution is used to change the number of output channels to Nc×Na.

[0074] The task of the regression branch is to output the target position parameters and category confidence scores, and feed the obtained target position parameters and category confidence scores back to the network to guide network training. The specific method is to send the input features to the CBL layer to obtain the target position information and realize feature normalization. Then, CBL is used again to divide the position parameter feature map and the confidence score feature map, and the number of output channels is changed to 4×Na and 1×Na respectively through 1×1 convolution operations.

[0075] Step 7: Perform non-maximum suppression (NMS) on the target category and location information (location parameters and confidence scores) output by ECA-DDH to obtain the final output result of the network.

[0076] The present embodiment is further described below by experiments:

[0077] The deep learning framework used in the ST-CFE network is Pytorch1.8 and Cuda11.1. The experiments are all conducted under Linux 4.15.0 system, and the hardware configuration used is: Intel Xeon Gold5120 CPU@2.20GHz, NVIDIA GeForce RTX3080 GPU (10G memory). In the image preprocessing stage, 100% mosaic enhancement is used; the total loss function is optimized and trained using the adam optimizer, with a momentum parameter of 0.937; the initial learning rate is 0.01; the weight decay is 0.0005; the number of weight warm-up rounds is 3 (a larger learning rate is used in the first 3 epochs); the batch size is set to 16.

[0078] Table 1 shows that the performance of this embodiment on the TT100K dataset exceeds that of the mainstream object detection network YOLOv7, and its mAP@0.5 is improved by 6.4% over YOLOv5 and 1.9% over YOLOv7.

[0079] Table 1 shows that the method of this embodiment shows the best results in mAP@0.5 and mAP@0.5:0.95 on the self-built high-bay library dataset, which are 92.0% and 48.6% respectively, which are 4.9% and 5.2% higher than YOLOv5, and 1% and 4.4% higher than the mainstream detection network YOLOv7. This shows that the method of this embodiment has high performance in accuracy, comprehensiveness of small target detection and small target positioning.

[0080] Table 1 Performance comparison results of the method in this embodiment and other algorithms

[0081]

[0082] Table 2 shows the performance improvement of the method in this embodiment compared with other networks in terms of mAP@0.5 for various target detection indicators. The performance of the extremely small targets, such as hard hats and dropped cigarette boxes, is improved by 6.9% and 5.5% respectively compared with YOLOv5, and by 2.4% and 1.4% compared with the mainstream detection network YOLOv7. This verifies the superiority of the method in this embodiment for small target detection in the high-bay warehouse scenario.

[0083] Table 2 Comparative experimental results of mAP@0.5 / % for each target in the self-built high-rise library dataset

[0084]

[0085] Example 2

[0086] A target detection method based on small target contrast feature enhancement (ST-CFE) includes the following 7 steps:

[0087] Step 1. Build the ST-CFE network based on small target contrast feature enhancement proposed in this embodiment, including the backbone network CSPDarknet53, the fast spatial pyramid pooling cross-layer fusion module (SPPF-CLFM), the bidirectional feature pyramid network (BiFPN), the small target contrast feature enhancement module (ST-CFEM) and the efficient channel attention decoupling detection head (ECA-DDH);

[0088] Step 2. The scene image is input into the backbone network CSPDarknet53 to obtain feature maps of 4 different scales;

[0089] Step 3. Input the last feature map of the backbone network into the fast spatial pyramid pooling cross-layer fusion module (SPPF-CLFM) to obtain information of small objects at multiple scales and reduce the loss of detail information in the spatial dimension;

[0090] Step 4. Input the first three feature maps of different scales of the backbone network and the output feature map of SPPF-CLFM into BiFPN to avoid the loss of feature information and retain the detailed information of small target features;

[0091] Step 5. Input the feature map in BiFPN into ST-CFEM to weaken the large target information in the shallow feature map and enhance the expression of small target feature information;

[0092] Step 6. Input the outputs of the two ST-CFEMs and the feature maps of the last two layers of BiFPN into ECA-DDH, so that the networks can learn the semantic information and location information of the target without interfering with each other, thus improving the classification accuracy;

[0093] Step 7. Input the outputs of the four ECA-DDHs into NMS to obtain the final output result of the network.

[0094] In this embodiment, the ST-CFE target detection method in step 1 fully utilizes the characteristics of high-resolution feature maps with more position and detail information by adding a small target detection head to the shallow feature map C2; introduces SPPF-CLFM to reduce information loss in the spatial dimension; builds a BiFPN structure to promote multi-scale feature fusion; introduces ST-CFEM to weaken the large target information in the shallow feature map, contrasts and enhances the expression of small target feature information, and improves the network's detection performance for small targets; introduces ECA-DDH to enable the network to learn the target's detail information and position information without interfering with each other, and adaptively learn the importance of each channel without losing feature information.

[0095] In step 3, the fast spatial pyramid pooling cross-layer fusion module (SPPF-CLFM) designed in this embodiment uses continuous maximum pooling and channel splicing to capture target information at different scales, and uses skip connections and pixel-by-pixel addition to reduce the loss of target information in the spatial dimension.

[0096] In step 5, the small target contrast feature enhancement module (ST-CFEM) designed in this embodiment uses two-way parallel dilated convolutions of different sizes to obtain different receptive fields, uses channel attention to obtain more accurate feature information of the feature map in the channel dimension, and uses the shallow feature map to subtract the deep feature map to weaken the large target information in the shallow feature, thereby contrasting and enhancing the expression of small target feature information.

[0097] In step 6, the efficient channel attention decoupled detection head (ECA-DDH) designed in this embodiment decouples the detection head into classification and regression branches to better learn the semantic information and location information of the target. In the classification branch, efficient channel attention (ECA) is introduced to adaptively learn the importance of each channel without losing feature information.

[0098] This embodiment proposes a small target contrast feature enhancement (ST-CFE) network, which is improved on the basis of YOLOv5 and adds a fast spatial pyramid pooling cross-layer fusion module (Spatial Pyramid Pooling Fast-Cross Layer Fusion Module, SPPF-CLFM), a bidirectional feature pyramid network (Bi-directional Feature Pyramid Network, BiFPN), a small target contrast feature enhancement module (Small Target Contrast Feature Enhancement Module, ST-CFEM) and an efficient channel attention decoupled detection head (Efficient Channel Attention Decoupled Detection Head, ECA-DDH).

[0099] Specifically, the method of this embodiment includes a backbone network, a neck network and a detection head. The backbone network uses CSPDarknet53 (Cross Stage Partial Darknet 53) for feature extraction. A fast spatial pyramid pooling cross-layer fusion module (SPPF-CLFM) is added to the neck network, and the maximum pooling and channel splicing are used to capture small target information at different scales, and the skip connection and pixel-by-pixel addition operation are used to reduce the loss of target information in the spatial dimension. A small target detection head is added to the shallow feature map with high resolution and more position and detail information, and the bidirectional feature pyramid network BiFPN is used to avoid the loss of feature information and retain the detail information of small target features. However, the large target information in the shallow feature map (C2-C3) will interfere with the extraction of small target information. For this reason, a small target contrast feature enhancement module (ST-CFEM) is constructed, and two-way parallel hole convolutions of different sizes are used to obtain different receptive fields. Channel attention is used to obtain more accurate feature information of the feature map in the channel dimension, and the shallow feature map is subtracted from the deep feature map to weaken the large target information contrast in the shallow feature map to enhance the expression of small target feature information, thereby improving the network's detection performance for small targets. An efficient channel attention decoupled detection head (ECA-DDH) is constructed to decouple the original coupled detection head into classification and regression branches, so that the network can learn the details and location information of the target without interfering with each other. The classification branch adds efficient channel attention (ECA) so that the network can adaptively learn the importance of each channel without losing feature information. Finally, the output of ECA-DDH is subjected to non-maximum suppression (NMS) to obtain the final output of the network.

[0100] The above-mentioned embodiments only express the specific implementation of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A target detection method based on small target contrast feature enhancement, characterized in that: The following steps are involved: Step 1: Build a ST-CFE network based on small target contrast feature enhancement; Step 2: Input the image into the backbone network CSPDarknet53 of the small target contrast feature enhancement ST-CFE network to obtain feature maps of different scales; Step 3: Input the last feature map of the backbone network into the fast spatial pyramid pooling cross-layer fusion module SPPF-CLFM, and use the fast spatial pyramid pooling cross-layer fusion module SPPF-CLFM to capture the target information at multiple scales and reduce the target information loss in the spatial dimension; Step 4: Add a small target detection head to the shallow feature map with high resolution and more position and detail information, input the shallow feature map of the backbone network and the output feature map of the fast spatial pyramid pooling cross-layer fusion module SPPF-CLFM into the bidirectional feature pyramid network BiFPN, and use the bidirectional feature pyramid network BiFPN to promote multi-scale feature fusion, avoid the loss of feature information, and retain the detail information of small target features; Step 5: Input the feature map in the bidirectional feature pyramid network BiFPN into the small target contrast feature enhancement module ST-CFEM. Through the small target contrast feature enhancement module ST-CFEM, the large target information in the shallow feature map is weakened, and the small target feature expression is contrast enhanced; Step 6: Build an efficient channel attention decoupling detection head ECA-DDH, and input the output of the small target contrast feature enhancement module ST-CFEM and the last two layers of feature maps of the bidirectional feature pyramid network BiFPN into the efficient channel attention decoupling detection head ECA-DDH; Step 7: Perform non-maximum suppression (NMS) on the target category and location information output by the efficient channel attention decoupling detection head ECA-DDH to obtain the final output result.

2. The target detection method based on small target contrast feature enhancement according to claim 1, characterized in that: In step 3, the fast spatial pyramid pooling cross-layer fusion module SPPF-CLFM uses continuous maximum pooling and channel splicing to capture target information at different scales, and uses skip connections and pixel-by-pixel addition to reduce the loss of target information in the spatial dimension.

3. The target detection method based on small target contrast feature enhancement according to claim 1, characterized in that: The step 5 is specifically as follows: For the shallow feature map, two-way parallel dilated convolution is used to obtain different receptive fields, and channel attention is used to learn the importance of different channels. Finally, the deep feature map is subtracted from the shallow feature map after feature enhancement to weaken the large target information of the shallow feature map and enhance the expression of small target feature information.

4. The target detection method based on small target contrast feature enhancement according to claim 1, characterized in that: In step 6, by decoupling the coupled detection head into classification and regression branches, the classification and regression branches are allowed to use different loss functions and weights to better balance the training process of the two branches; the classification and regression branches are allowed to focus on the target semantic information and position information respectively, reducing mutual interference between branches and improving the network's ability to classify and locate targets; efficient channel attention ECA is added to the classification branch to adaptively learn the importance of each channel without losing feature information, thereby improving the network classification accuracy.

Images