Target detection method and device based on anti-interference characteristic information compensation

Abstract

The invention relates to the technical field of target detection, in particular to a target detection method, device and equipment based on anti-interference feature information compensation and a computer storage medium. According to the target detection method based on anti-interference feature information compensation, detail feature information on a lower-layer feature map is fused in a cross-hierarchy manner through a dynamic position aggregation network, so that small target feature information is effectively compensated, and cross-hierarchy information fusion can meet a semantic gap; and a dynamic up-sampling operator is introduced to effectively suppress cross-layer high-magnification up-sampling noise, so that the precision of small target detection is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of object detection, and in particular to an object detection method, device, equipment and computer storage medium based on anti-interference feature information compensation. Background Technique

[0002] In the field of computer vision, small object detection is an important and challenging task, which is widely used in many fields such as video surveillance, autonomous driving, and remote sensing image analysis. Small objects usually refer to objects with small sizes and few pixels in an image, and their detection accuracy and recall rate are relatively low. Traditional object detection methods often cannot effectively detect these small objects. In recent years, the rapid development of deep learning, especially convolutional neural networks (CNNs), has greatly promoted the progress of object detection technology. However, although significant progress has been made in medium and large object detection, small object detection still faces many difficulties, mainly due to the easy loss of small object feature information, susceptibility to noise interference, and sensitivity to bounding box perturbations. Therefore, how to improve the accuracy of small object detection has become a key issue in the field of computer vision.

[0003] The mainstream compensation of small object feature information is achieved through feature information fusion. It mainly includes two types of solutions: (1) Small object detection is a challenging task in the field of object detection, which has attracted a large amount of research and made significant progress. Two key problems in small object detection are feature information loss and noise interference. Feature Pyramid Network (FPN) and its variants have made significant progress in the field of small object detection. However, traditional FPN still faces challenges when dealing with small objects. This is mainly due to the nature that high-level features are rich in semantics but low in resolution, and low-level features are high in resolution but weak in semantics, which will cause small object features to be easily submerged by background noise. To solve these problems, researchers have proposed various improved FPN structures. For example: 1. High-Frequency Spatial Awareness Feature Pyramid Network (HS-FPN), which uses a high-pass filter to generate high-frequency responses to enhance small object features, and adds a spatial dependence perception module to capture spatial dependence and improve detection performance; 2. Asymptotic Feature Pyramid Network (AFPN), which gradually fuses features of non-adjacent levels to avoid semantic gaps, and combines adaptive spatial fusion operations to alleviate conflicts between multi-object information, thereby improving detector performance; 3. Enhanced FPN (AugFPN), which improves multi-scale feature learning through mechanisms such as consistency supervision, residual feature enhancement, and soft RoI selection, thereby improving object detection performance; 4. Extended Feature Pyramid Network (EFPN), which has an ultra-high-resolution pyramid designed specifically for small object detection. This network effectively extracts reliable regional detail information and spreads detail perception ability at different scales.

[0004] (2) In small target detection tasks, due to the small size, weak features, and easy to be submerged by the background of the targets, traditional bounding box regression loss functions (such as GIoU, DIoU, CIoU) are often not sensitive enough to the localization of small targets during the optimization process. Therefore, researchers have proposed various loss function designs for small targets to improve the performance of detectors for small targets. A typical method to enhance the localization performance is to improve the IoU-based loss function. For example, the α-IoU loss introduces an exponential decay term, making the loss more sensitive to small targets in the low IoU range. Another type of method introduces distribution-based metric concepts, such as the normalized Wasserstein distance (NWD) and receptive field label assignment based on Gaussian (RFLA). This method characterizes the geometric relationship and spatial structure between them by measuring the distribution difference between the predicted box and the ground truth box. This method is particularly robust and effective for smaller targets. All in all, the loss function design for small target detection is evolving towards higher perceptual accuracy and stronger geometric robustness. A reasonable loss function can not only enhance the model's regression ability for small targets but also alleviate the problem of imbalance between positive and negative samples to a certain extent.

[0005] In the first type of solution, when compensating for feature information by the top-down information interaction feature fusion method of consecutive levels, there is a certain semantic gap between consecutive level features. This semantic gap between consecutive levels will cause a decrease in the detection accuracy of small targets and at the same time lack the design of relevant modules for perceiving the subtle features of small targets.

[0006] In the second type of solution, a series of improved designs of IOU and distribution-based metrics can alleviate the problem of bounding box perturbation in small target detection to a certain extent. However, both design solutions have certain design defects. The IoU measurement method is not accurate enough for the regression loss of small targets, while the relevant designs of NWD and RFLA will lose the effectiveness of the measurement for medium and large target detection. Therefore, both of these solutions have limitations at the algorithm design level.

[0007] As mentioned above, the current mainstream method for compensating feature information in small target detection uses the consecutive level information fusion method, which will cause a semantic gap between high and low level feature maps and at the same time ignores the interference of sampling noise and bounding box perturbation on the feature representation of small targets, resulting in low accuracy of small target detection. Summary of the Invention

[0008] Therefore, the technical problem to be solved by the present invention is to overcome the problem of low accuracy in small target detection in the prior art.

[0009] To solve the above technical problem, the present invention provides a target detection method based on anti-interference feature information compensation, including: Obtain the image to be detected; Input the image to be detected into the backbone network for multi-level feature extraction to obtain feature maps of multiple different levels; Input the feature maps of the multiple different levels into the dynamic position aggregation network, perform a dynamic upsampling operation based on the highest-level feature map, and cross-level fuse to introduce the low-level feature maps to obtain a small target detection feature map; Input the small target detection feature map into the detection head network to obtain a small target detection result.

[0010] Preferably, the step of inputting the feature maps of the multiple different levels into the dynamic position aggregation network, performing a dynamic upsampling operation based on the highest-level feature map, and cross-level fusing to introduce the low-level feature maps to obtain a small target detection feature map includes: Input the feature maps of the multiple different levels into the dynamic position aggregation network. The processing process of the dynamic position aggregation network includes performing a multiple dynamic upsampling operation on the highest-level feature map, then fusing and introducing the feature map corresponding to the current level to obtain the fused feature map of the current level, using the fused feature map of the current level as the input feature map for the next multiple dynamic upsampling operation, repeating this step multiple times, and using the fused feature map of the lowest level as the small target detection feature map.

[0011] Preferably, before using the fused feature map of the current level as the input feature map for the next multiple dynamic upsampling operation, it further includes: Input the fused feature map of the current level into the C2FM module for processing. The processing process of the C2FM module includes: After inputting the fused feature map of the current level into the CBM module for processing, it is successively processed through the segmentation module and multiple bottleneck layers, and the output of the segmentation module and the output of each bottleneck layer are concatenated and fused, and then input into the CBM module for processing again to obtain the output of the C2FM module. The CBM module includes a batch normalization layer, a Mish activation function layer, and a convolution operation layer.

[0012] Preferably, the specific steps of the dynamic upsampling operation include: Input the input feature map into two linear transformation layers respectively to obtain a first feature map and a second feature map; Input the first feature map into the sigmoid function layer for processing, multiply by a specified coefficient, and then perform a dot product operation with the second feature map to output an offset; Based on the offset and the original sampling grid, obtain a sampling set; Based on the sampling set, through grid sampling operation, output a dynamic sampling result feature map.

[0013] Preferably, after obtaining the small target detection feature map, it further includes: After inputting the small object detection feature map into the CBS module for processing, it is fused with the fused feature map at the nth level and then input into the C2FM module for processing to obtain the medium object detection feature map. The CBS module includes a batch normalization layer, a SiLU activation function layer, and a convolution operation layer, and n is a preset constant.

[0014] Preferably, after obtaining the small object detection feature map, it further includes: After inputting the small object detection feature map into the CBS module for processing, it is fused with the fused feature map at the nth level, and after passing through the SCDown layer, it is fused with the highest-level feature map. Finally, it is input into the C2FM module for processing to obtain the large object detection feature map.

[0015] Preferably, the detection head network introduces a scale-sensitive loss function to detect the small object detection feature map. Among them, the scale-sensitive loss function is obtained based on the CIoU loss function, the NWD loss function, and a scale factor, and the scale factor is determined based on the proportion of the small object in the image to be detected.

[0016] The present invention also provides an object detection device based on anti-interference feature information compensation, including: An image acquisition module for acquiring the image to be detected; A feature extraction module for inputting the image to be detected into the backbone network for multi-level feature extraction to obtain multiple feature maps at different levels; A feature fusion module for inputting the multiple feature maps at different levels into the dynamic position aggregation network. The processing process of the dynamic position aggregation network includes: performing a multi-fold dynamic upsampling operation on the highest-level feature map, fusing and introducing the corresponding feature map of the current level to obtain the fused feature map of the current level, inputting the fused feature map of the current level into the C2FM module for processing and using it as the input feature map for the next multi-fold dynamic upsampling operation, repeating this step multiple times, and using the fused feature map of the lowest level as the small object detection feature map. The processing process of the C2FM module includes: After inputting the fused feature map of the current level into the CBM module for processing, it is successively processed through a segmentation module and multiple bottleneck layers, and the output of the segmentation module and the output of each bottleneck layer are spliced and fused, and then input into the CBM module for processing again to obtain the output of the C2FM module. The CBM module includes a batch normalization layer, a Mish activation function layer, and a convolution operation layer; An object detection module for inputting the small object detection feature map into the detection head network to obtain the small object detection result.

[0017] The present invention also provides an object detection device based on anti-interference feature information compensation, including: A memory for storing a computer program; A processor for implementing the steps of the above-mentioned object detection method based on anti-interference feature information compensation when executing the computer program.

[0018] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned object detection method based on anti-interference feature information compensation are implemented.

[0019] The above technical solution of the present invention has the following advantages compared with the prior art: In the object detection method based on anti-interference feature information compensation of the present invention, detailed feature information on a lower-level feature map is incorporated across levels through a dynamic position aggregation network, so as to effectively compensate for small target feature information, and the cross-level information fusion can bridge the semantic gap. By introducing a dynamic upsampling operator, the cross-level high magnification upsampling noise is effectively suppressed, and the accuracy of small target detection is effectively improved. Description of the Drawings

[0020] In order to make the content of the present invention easier to be clearly understood, the following further details the present invention according to specific embodiments of the present invention in conjunction with the drawings, where: Figure 1 is a flowchart of the implementation of an object detection method based on anti-interference feature information compensation provided by the present invention; Figure 2 is an overall architecture diagram of an object detection method based on anti-interference feature information compensation in a specific embodiment of the present invention; Figure 3 is a structural block diagram of an object detection device based on anti-interference feature information compensation provided by an embodiment of the present invention. Detailed Embodiments

[0021] The core of the present invention is to provide an object detection method, device, equipment and computer storage medium based on anti-interference feature information compensation, which effectively improves the accuracy of small target detection.

[0022] In order to enable those skilled in the art to better understand the solution of the present invention, the following further details the present invention in conjunction with the drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0023] The small object detection methods in the prior art mainly use feature pyramids for feature information fusion, which easily leads to semantic gaps between levels and affects the accuracy of small object detection. Due to the relatively low pixel proportion of small objects, they are easily affected by bounding box perturbations, posing challenges to small object detection; for example, in the drone target detection task, the proportion of small objects is relatively high. Due to limited pixel quantity and insufficient feature information, compared with medium and large objects containing more detailed visual semantics, small objects suffer more severe feature loss during the downsampling process. After multiple downsampling operations, small objects may be compressed to pixel-level or even sub-pixel-level representations, making their detection in high-level feature maps extremely challenging. To solve this problem, the present invention proposes a Dynamic Position Aggregation (DPA) network, which explicitly enhances the target area in the shallow and deep layers, thereby enabling full utilization of the complementary saliency representations of small objects. Specifically, DPA extracts effective feature information from low-level feature maps, thereby reducing the negative impact of feature loss during the feature extraction process on small object detection. In addition, the cross-level feature fusion architecture of the DPA network effectively avoids the semantic gap caused by continuous hierarchical feature fusion. By integrating cross-level information, the DPA network introduces low-level detailed features from low-level feature maps to provide additional guidance for small object detection, thereby promoting the effective fusion of small object spatial and position information. By organically combining high-level global context information with low-level local detailed information, DPA further enhances the small object detection ability of the network.

[0024] Please refer to FIG. 1, Figure 1 which is a flowchart for implementing an object detection method based on anti-interference feature information compensation provided by the present invention; the specific operation steps are as follows: S101: Obtain the image to be detected; S102: Input the image to be detected into the backbone network for multi-level feature extraction to obtain multiple feature maps of different levels; S103: Input the multiple feature maps of different levels into the dynamic position aggregation network. The processing process of the dynamic position aggregation network includes performing a multiple dynamic upsampling operation on the highest-level feature map, then fusing and introducing the corresponding feature map of the current level to obtain the fused feature map of the current level, using the fused feature map of the current level as the input feature map for the next multiple dynamic upsampling operation, repeating this step multiple times, and using the fused feature map of the lowest level as the small object detection feature map; In a specific embodiment, the highest-level feature map of is subjected to a 4-fold dynamic upsampling operation, and then fused and introduced with the corresponding feature map of to obtain the fused feature map of ; The fused feature map of the level As the input feature map for the next 2x dynamic upsampling operation and fuse and introduce The feature map corresponding to the level , obtain The fused feature map of the level , and The fused feature map of the level As the small target detection feature map .

[0025] Based on the above embodiments, before using the fused feature map of the current level as the input feature map for the next multi-fold dynamic upsampling operation, it further includes: Input the fused feature map of the current level into the C2FM module for processing; In another specific embodiment, The highest level feature map of the level After performing a 4x dynamic upsampling operation, fuse and introduce The feature map corresponding to the level , obtain The fused feature map of the level ; Use the fused feature map of the level after being processed by the C2FM module As the input feature map for the next 2x dynamic upsampling operation and fuse and introduce The feature map corresponding to the level , obtain The fused feature map of the level , use the fused feature map of the level after being processed by the C2FM module As the small target detection feature map .

[0026] Among them, the processing process of the C2FM module includes: Input the fused feature map of the current level into the CBM module for processing, then sequentially pass through the segmentation module and multiple bottleneck layers for processing, and splice and fuse the output of the segmentation module and the output of each bottleneck layer, and then input it into the CBM module for processing again to obtain the output of the C2FM module, Among them, the CBM module includes a batch normalization layer, a Mish activation function layer, and a convolution operation layer.

[0027] Based on the above embodiments, after obtaining the small target detection feature map, it further includes: Input the small target detection feature map into the CBS module for processing, fuse it with the fused feature map of the nth level and then input it into the C2FM module for processing to obtain the medium target detection feature map, The CBS module includes a batch normalization layer, a SiLU activation function layer, and a convolution operation layer, where n is a preset constant.

[0028] S104: Input the small target detection feature map into the detection head network to obtain the small target detection result.

[0029] Based on the above embodiments, this embodiment elaborates on step S102 in detail: Input the image to be detected into the backbone network for multi-level feature extraction to obtain N feature maps at different levels ( ) In one embodiment, the specific method of feature extraction includes downsampling operations. In a specific embodiment, N = 5.

[0030] Based on the above embodiments, this embodiment elaborates on step S103 in detail: In the DPA network, high-magnification upsampling is required to match the size of the low-level feature map. However, traditional nearest-neighbor interpolation and bilinear interpolation introduce additional sampling noise, interfering with the feature representation of small targets. To reduce the interference caused by high-magnification upsampling noise, the present invention uses dynamic upsampling instead of traditional nearest-neighbor interpolation and bilinear interpolation in the DPA network; specifically: The specific steps of the dynamic upsampling operation include: Input the input feature map into two linear transformation layers and respectively to obtain a first feature map and a second feature map ; Input the first feature map into the sigmoid function layer for processing and multiply it by a specified coefficient k, then perform a dot product operation with the second feature map to output an offset ; In a specific embodiment, k = 0.5; Based on the offset and the original sampling grid to obtain a sampling set; In one embodiment, the sampling set is obtained by adding the offset and the original sampling grid ; Based on the sampling set, through grid sampling operations, output the dynamic sampling result feature map .

[0031] In a specific embodiment, n = 3. After inputting the small target detection feature map into the CBS module for processing, is obtained, and it is fused with the feature map at the After being processed by the C2FM module through fusion, the medium target detection feature map is obtained ; In another specific embodiment, n = 3. After the small target detection feature map is input into the CBS module for processing, is obtained, and it is fused with the fusion feature map of the nth level After being processed by the C2FM module through fusion, the medium target detection feature map is obtained ; Among them, the processing steps of the C2FM module are the same as those above and will not be elaborated here.

[0032] Based on the above embodiments, after obtaining the small target detection feature map, it further includes: The small target detection feature map is input into the CBS module for processing, then fused with the fusion feature map of the nth level, and after being processed by the SCDown layer, it is fused with the highest level feature map. Finally, it is input into the C2FM module for processing to obtain the large target detection feature map; Among them, the spatial-channel decoupled downsampling (SCDown) layer reduces the dimension of the feature map by decoupling the spatial and channel information, while retaining important feature information.

[0033] In a specific embodiment, n = 3. After the small target detection feature map is input into the CBS module for processing, is obtained, and it is fused with the fusion feature map of the nth level After being processed by the SCDown layer after fusion, is obtained, and it is fused with the highest level feature map Finally, it is input into the C2FM module for processing to obtain the large target detection feature map ; In another specific embodiment, n = 3. After the small target detection feature map is input into the CBS module for processing, is obtained, and it is fused with the fusion feature map of the nth level After being processed by the SCDown layer after fusion, is obtained, and it is fused with the highest level feature map Finally, it is input into the C2FM module for processing to obtain the large target detection feature map .

[0034] In this embodiment, the image to be detected is input into the backbone network for feature extraction, and then feature information fusion is performed in the designed DPA module, effectively combining low-level detailed feature information and high-level context semantic information to effectively compensate for small target feature information, avoiding the reduction of small target detection accuracy caused by the semantic gap brought by the fusion of continuous hierarchical feature information; the upsampling part uses dynamic upsampling for 4x and 2x upsampling. Dynamic upsampling has a better sampling noise suppression effect compared to traditional nearest neighbor interpolation and bilinear interpolation; the C2FM module proposed in DPA can effectively improve the network's perception ability of small target fine feature information.

[0035] Based on the above embodiments, this embodiment details step S104: Input the small target detection feature map into the detection head network (Delect module) to obtain the small target detection result; Input the medium target detection feature map into the detection head network to obtain the medium target detection result; Input the large target detection feature map into the detection head network to obtain the large target detection result.

[0036] The detection head network of the present invention introduces a scale-sensitive loss function to detect the small target detection feature map. Among them, the scale-sensitive loss function is obtained based on the CIoU loss function, the NWD loss function, and a scale factor, and the scale factor is determined based on the proportion of the small target in the image to be detected.

[0037] Although the IoU loss function performs well for medium and large targets, small targets are more sensitive to bounding box perturbations. Therefore, the IoU metric is not suitable for small target regression. In traditional scenarios dominated by medium and large targets, this problem has little impact on the performance of the detector. However, in scenarios such as drone aerial photography images with a relatively large proportion of small targets, bounding box perturbations will significantly reduce the detection performance of the model. To solve this problem, the present invention designs SSL to measure the regression loss of targets of different scales. SSL is a combined regression loss function that combines the advantages of the CIoU and NWD metrics. SSL introduces a scale factor λ to adjust the relative contributions between the two according to the proportion of small targets. This design enables SSL to effectively increase the proportion of small target positive samples during training, enabling the model to more effectively learn small target features without reducing the feature learning ability for medium and large targets, thereby improving the overall small target detection performance.

[0038] According to the definition of the Wasserstein distance in optimal transport theory, the bounding box and can be modeled as Gaussian distributions and , the distance between these two distributions can be expressed as:

[0039] where represents the F-norm.

[0040] Since has a value range of , it cannot be directly used for label assignment. The value range of IoU is . To align its value range with that of IoU, the present invention selects to use an exponential non-linear transformation function to map the Wasserstein distance between Gaussian distributions to another space, so that its value range becomes . The normalized Wasserstein distance is shown as follows:

[0041] where is a constant.

[0042] Therefore, the loss function based on NWD can be expressed as:

[0043] CIoU is defined as follows:

[0044] where is defined as:

[0045] where is defined as follows:

[0046] The regression loss function of the present invention is defined as follows:

[0047] where is a scale perception factor, which can be adjusted according to the proportion relationship between small targets and medium and large targets, and its value range is , and is used to adjust the weight between CIoU and NWD losses.

[0048] Since small target detection is easily affected by the interference caused by bounding box perturbations. Therefore, in order to reduce the interference of bounding box perturbations on small target detection and at the same time ensure that the measurement of medium and large targets is not negatively affected, the present invention designs an SSL loss function for target regression loss measurement, effectively reducing the influence of bounding box perturbations on small target detection.

[0049] As Figure 2 ,Figure 2 This is the overall architecture diagram of the object detection method based on anti-interference feature information compensation in a specific embodiment of the present invention. The image to be detected is input into the backbone network for feature extraction, and then feature information fusion is performed in the designed DPA network module. By effectively combining low-level detailed feature information and high-level context semantic information, the feature information of small objects is effectively compensated, avoiding the reduction of small object detection accuracy caused by the semantic gap brought by continuous hierarchical feature information fusion. The upsampling part uses dynamic upsampling for 4-fold and 2-fold upsampling. Dynamic upsampling has a better sampling noise suppression effect compared with traditional nearest neighbor interpolation and bilinear interpolation. The C2FM module proposed in DPA can effectively improve the network's perception ability of the subtle feature information of small objects. In the regression part of the detection head, a regression loss metric function called SSL is designed. This loss function can effectively combine the advantages of CIoU loss and NWD loss, and can effectively reduce the impact of bounding box perturbation on small object detection without affecting the detection effect of medium and large objects.

[0050] Please refer to Figure 3 , Figure 3 This is the structural block diagram of an object detection device based on anti-interference feature information compensation provided by an embodiment of the present invention. The specific device may include: An image acquisition module 100 for acquiring the image to be detected; A feature extraction module 200 for inputting the image to be detected into the backbone network for multi-level feature extraction to obtain feature maps of multiple different levels; A feature fusion module 300 for inputting the feature maps of multiple different levels into the dynamic position aggregation network. In one embodiment, the processing process of the dynamic position aggregation network includes: performing multi-fold dynamic upsampling operation on the highest-level feature map, fusing and introducing the corresponding feature map of the current level to obtain the fused feature map of the current level, inputting the fused feature map of the current level into the C2FM module for processing, and using it as the input feature map for the next multi-fold dynamic upsampling operation. Repeat this step multiple times, and use the fused feature map of the lowest level as the small object detection feature map. The processing process of the C2FM module includes: After inputting the fused feature map of the current level into the CBM module for processing, it is successively processed by the segmentation module and multiple bottleneck layers, and the output of the segmentation module and the output of each bottleneck layer are spliced and fused, and then input into the CBM module for processing again to obtain the output of the C2FM module. The CBM module includes a batch normalization layer, a Mish activation function layer, and a convolution operation layer; An object detection module 400 for inputting the small object detection feature map into the detection head network to obtain the small object detection result.

[0051] The object detection device of this embodiment is used to implement the aforementioned object detection method. Therefore, the specific implementation manners in the object detection device can be seen in the embodiment part of the previous object detection method. For example, the image acquisition module 100, the feature extraction module 200, the feature fusion module 300, and the object detection module 400 are respectively used to implement steps S101, S102, S103, and S104 in the above object detection method. Therefore, the specific implementation manners can refer to the descriptions of the corresponding individual part embodiments and will not be elaborated here.

[0052] A specific embodiment of the present invention also provides an object detection device, including: a memory for storing a computer program; a processor for implementing the steps of the above object detection method when executing the computer program.

[0053] A specific embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and the computer program implements the steps of the above object detection method when executed by a processor.

[0054] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0055] This application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0056] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing devices to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured product including an instruction device, and the instruction device implements the functions in the flow Figure 1one process or multiple processes and / or blocks Figure 1 the functions specified in one block or multiple blocks.

[0057] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 the functions specified in one block or multiple blocks.

[0058] Obviously, the above embodiments are only examples given for clear illustration and are not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. And the obvious changes or modifications derived therefrom still fall within the protection scope of the present invention.

Claims

1. A target detection method based on anti-interference feature information compensation, characterized in that, Including: Obtain the image to be detected; Input the image to be detected into the backbone network for multi-level feature extraction to obtain multiple feature maps of different levels; Input the multiple feature maps of different levels into the dynamic position aggregation network. The processing process of the dynamic position aggregation network includes: after performing multi-fold dynamic upsampling on the highest-level feature map, fusing and introducing the feature map corresponding to the current level to obtain the fused feature map of the current level, inputting the fused feature map of the current level into the C2FM module for processing and using it as the input feature map for the next multi-fold dynamic upsampling operation, repeating this step multiple times, and using the fused feature map of the lowest level as the small object detection feature map. The processing process of the C2FM module includes: Input the fused feature map of the current level into the CBM module for processing, then successively pass through the segmentation module and multiple bottleneck layers for processing, splice and fuse the output of the segmentation module and the output of each bottleneck layer, and then input it into the CBM module for processing again to obtain the output of the C2FM module. The CBM module includes a batch normalization layer, a Mish activation function layer, and a convolution operation layer; Input the small object detection feature map into the detection head network to obtain the small object detection result.

2. The object detection method based on anti-interference feature information compensation according to claim 1, wherein The specific steps of the dynamic upsampling operation include: Input the input feature map into two linear transformation layers respectively to obtain the first feature map and the second feature map; Input the first feature map into the sigmoid function layer for processing, multiply it by a specified coefficient, and then perform a dot product operation with the second feature map to output the offset; Based on the offset and the original sampling grid, obtain the sampling set; Based on the sampling set, through grid sampling operation, output the dynamic sampling result feature map.

3. The target detection method based on anti-interference feature information compensation according to claim 1, wherein, After obtaining the small object detection feature map, it further includes: Input the small object detection feature map into the CBS module for processing, fuse it with the fused feature map of the nth level, and then input it into the C2FM module for processing to obtain the medium object detection feature map. The CBS module includes a batch normalization layer, a SiLU activation function layer, and a convolution operation layer, where n is a preset constant.

4. The target detection method based on anti-interference feature information compensation according to claim 3, wherein, After obtaining the small object detection feature map, it further includes: Input the small object detection feature map into the CBS module for processing, fuse it with the fused feature map of the nth level, fuse it with the highest-level feature map after passing through the SCDown layer, and finally, input it into the C2FM module for processing to obtain the large object detection feature map.

5. The object detection method based on anti-interference feature information compensation according to claim 1, characterized in that The detection head network introduces a scale-sensitive loss function to detect the small object detection feature map. Among them, the scale-sensitive loss function is obtained based on the CIoU loss function, the NWD loss function, and a scale factor, and the scale factor is determined based on the proportion of the small object in the image to be detected.

6. An object detection device based on anti-interference feature information compensation, characterized in that, Including: An image acquisition module for obtaining the image to be detected; A feature extraction module for inputting the image to be detected into the backbone network for multi-level feature extraction to obtain multiple feature maps of different levels; A feature fusion module for inputting the feature maps of the multiple different levels into a dynamic position aggregation network. The processing process of the dynamic position aggregation network includes: performing a multi-fold dynamic upsampling operation on the highest-level feature map, fusing and introducing the feature map corresponding to the current level to obtain the fused feature map of the current level, inputting the fused feature map of the current level into a C2FM module for processing, and using it as the input feature map for the next multi-fold dynamic upsampling operation. Repeat this step multiple times, and use the fused feature map of the lowest level as the small target detection feature map. The processing process of the C2FM module includes: After inputting the fused feature map of the current level into a CBM module for processing, it is sequentially processed through a segmentation module and multiple bottleneck layers, and the output of the segmentation module and the output of each bottleneck layer are concatenated and fused, and then input into the CBM module for processing again to obtain the output of the C2FM module. The CBM module includes a batch normalization layer, a Mish activation function layer, and a convolution operation layer. A target detection module for inputting the small target detection feature map into a detection head network to obtain a small target detection result.

7. An object detection device based on anti-interference feature information compensation, characterized in that It includes: A memory for storing a computer program; A processor for implementing the steps of the target detection method based on anti-interference feature information compensation according to any one of claims 1 to 5 when executing the computer program.

8. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the steps of the target detection method based on anti-interference feature information compensation according to any one of claims 1 to 5 are implemented.

Images