A target detection method based on edge extraction and accurate positioning

Abstract

The present invention belongs to the field of target detection technology, and specifically relates to a target detection method based on edge extraction and accurate positioning. The EEALNet network model is constructed to set a feature extraction module, a coarse positioning module and a precise segmentation module. The feature extraction module extracts four features of different scales t1-t4, and the deepest layer features are sent to the coarse positioning module. The coarse positioning module fuses the deepest layer features with other hierarchical features to obtain a coarse positioning map. The precise segmentation module further refines the coarse prediction map by alternately paying attention to the target area and the target edge to obtain the final prediction map. In the feature extraction stage, a pyramid visual converter is used to perform feature extraction. In the coarse positioning stage, the features are roughly positioned by fusing them from deep to shallow using a mirror fusion method. In the precise segmentation stage, the highly similar foreground and background are accurately separated by alternately paying attention to the target itself and the target edge. The precise positioning of the camouflaged target is gradually achieved in stages.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and in particular relates to a target detection method based on edge extraction and accurate positioning. Background Art

[0002] The task of detecting camouflaged objects requires detecting them from images. This challenging task requires combating sophisticated camouflage strategies and distinguishing subtle differences in color, texture, and shape between foreground and background. With the development of neural networks, deep network-based camouflaged object detection technology has made it possible. Compared to human visual detection, deep network-based camouflaged object detection technology can quickly and accurately detect camouflaged objects, improving detection efficiency and accuracy while saving labor and time. Camouflaged object detection technology has broad application value in many fields.

[0003] There are currently two challenges in detecting camouflaged objects: the first is that the object is difficult to accurately locate due to its extreme similarity to the background in terms of color, texture, shape, etc. The second is that even if the object is accurately located, its blurred edges make it difficult to accurately segment it.

[0004] In recent years, in order to solve these two problems, some researchers have divided the camouflaged target detection process into two stages: search and recognition. Specifically, the search stage roughly locates the camouflaged target and obtains a rough positioning map, and the recognition stage refines the obtained rough positioning map to obtain a more accurate prediction map. Figure 1 As shown, the first row shows images with accurate positioning, and the second row shows images with inaccurate positioning. The impact of coarse positioning accuracy on the final results shows that the final results of images with inaccurate coarse positioning are lower than the average index of the entire dataset, while the final results of images with accurate coarse positioning are higher than the average index of the entire dataset.

[0005] Both SARNet and PFNet directly perform rough positioning on the deepest features extracted by the backbone network. However, the resolution of the deepest features extracted by the backbone network is too small, and most of the position information may be lost, resulting in inaccurate rough positioning of the camouflaged target. Then the subsequent segmentation process will definitely be inaccurate. For this reason, the FEDER network uses ASPP (atrous spatial convolutional pooling pyramid) to extract and locate the deepest features at multiple scales during the search phase. The positioning effect is improved, but there is still an inaccurate positioning problem for small camouflaged targets. At the same time, the edge information of the camouflaged target is extremely important for the camouflaged target detection task, so as much edge information as possible should be extracted. SARNet uses the dilation algorithm to extract the edge information of the camouflaged target, but the range of the extracted edge is limited.

[0006] The present invention proposes a new network EEALNet, which consists of a feature extraction stage, a coarse positioning stage, and a precise segmentation stage. It is intended to gradually achieve the precise positioning of camouflaged targets through multiple stages and multiple modules. Summary of the Invention

[0007] The purpose of the present invention is to provide a target detection method based on edge extraction and accurate positioning, so as to create a new network EEALNet. The network consists of a feature extraction stage, a coarse positioning stage, and a precise segmentation stage. It is intended to gradually achieve the precise positioning of camouflaged targets through multiple stages and multiple modules.

[0008] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows:

[0009] A target detection method based on edge extraction and accurate positioning includes the following steps:

[0010] S1: Constructing an EEALNet network model, wherein the EEALNet network model includes a feature extraction module, a coarse positioning module, and a precise segmentation module;

[0011] S2: The image is input into the EEALNet network model, and the feature extraction module extracts four features t1-t4 of different scales from the image;

[0012] S3: The deepest feature of the four features is sent to the coarse positioning module, which fuses the deepest feature with other hierarchical features, aligns the size of the s1 channel with the s4 channel, and reduces the number of s1 channels to 1, obtaining a coarse positioning map p1 and transmitting it to the precise segmentation module;

[0013] S4: The precise segmentation module further refines the rough prediction map by alternately paying attention to the target area and the target edge to obtain the final prediction map.

[0014] Preferably, the feature extraction module uses the PVTv2 network as the backbone network, and the input image is , extract four features of different scales t1-t4 in turn:

[0015] ;

[0016] Among them, I is the input image, R is a set of real numbers, H is the height, W is the width, and i is the number of layers extracted by the backbone network.

[0017] Preferably, the coarse positioning module includes an MFM module, a downsampling module and a U-ASPP module. The specific processing process of the image by the U-ASPP module is as follows:

[0018] S31: The deepest feature t4 enters the U-ASPP module and is sent to five different branches. Through 3×3 dilated convolutions with different expansion rates, the three branches with different receptive fields extract features at multiple scales to capture objects of different sizes.

[0019] S32: The remaining two branches obtain global information through pooling and 1×1 convolution. The three branches containing dilated convolution in step S31 add the features of the previous branch after the SE attention mechanism to the features of the current branch, and then send them together to the dilated convolution and SE attention mechanism for further feature extraction;

[0020] S33: Perform channel fusion on two adjacent branches in a bottom-up order, and finally perform channel fusion on the features of the five branches to output S4.

[0021] Preferably, the specific processing process of the MFM module is as follows:

[0022] S34: The deep fusion feature s i+1 Further extract features through 3×3 convolution;

[0023] S35: Use Pixshuffle in super-resolution task to upsample and resize to t i Aligned, after sigmoid activation function and t i Multiply to get m1, the specific formula is as follows:

[0024] ;

[0025] in, CBR (·) represents Conv+BN+Relu, Pixshuffle (·) denotes upsampling in super-resolution;

[0026] S36: Utilize s i+1 To guide t i Focus on important features, t i Aligned by downsampling module s i+1 The size of , after sigmoid activation function and s i+1 Multiply to get m2, the specific formula is as follows:

[0027] ;

[0028] in, DownSample (·) represents a 3×3 convolution with a stride of 2;

[0029] S37: After upsampling m2 through Pixshuffle and aligning it to the size of m1, perform channel fusion with m1, and then send it to the SE attention mechanism to focus on more important features and suppress irrelevant or redundant information to obtain s i ;

[0030] ;

[0031] in, SE (·) represents the SE attention mechanism, Cat (·) represents channel fusion.

[0032] Preferably, the precise segmentation module includes two EGRM modules and one FGRM module. The specific processing process of the EGRM module is as follows:

[0033] The ground truth map in the dataset is subtracted from the ground truth map using the dilation algorithm to obtain the edge map P3. At the same time, the true edge map P1 in the dataset is colored to obtain P2. P2 and P3 are fused together to obtain P4.

[0034] The rough prediction map p i+2 Extract the edge map E through dilation and erosion algorithms i+2 , the edge graph is combined with the shallow s i+1 Multiply the shallow features s i+1 It is guided and then sent to 3×3 convolution to further extract features, multiplied by the learnable parameter α, and the result is combined with f i+1 Or the results after s4 alignment size and number of channels are added and sent to CBR to further extract features and output the enhanced features f i , and finally f i Send it to the 3×3 convolution to change the number of channels to get the predicted image p i+1 , the specific formula is as follows:

[0035] ;

[0036] Where i=1 or 2, when i=2, f i+1 is s4, CBR(·) is Conv+BN+ReLU; C(·) represents channel fusion; Pixshuffle(·) represents upsampling in super-resolution; α It is a learnable parameter, initialized to 1, and Conv3 represents 3×3 convolution.

[0037] Preferably, the specific processing process of the FGRM module is as follows:

[0038] Directly use the rough prediction map p i+2 As input, edges are no longer extracted, and the rest of the feature processing is the same as the EGRM module. The specific formula is as follows:

[0039] ;

[0040] Among them, i=0, CBR(·) is Conv+BN+ReLU; C(·) represents channel fusion; Pixshuffle(·) represents upsampling in super-resolution; α is a learnable parameter initialized to 1, and Conv3 represents 3×3 convolution.

[0041] Preferably, the process of calculating the loss by setting a loss function is also included, as follows:

[0042] will p i , i=1.2.3.4, adjust to the same size as the true value image and calculate the loss, p1 is the final output image, using binary cross entropy loss L bce Sum intersection loss L iou As a loss function. Binary cross entropy loss L bce Sum intersection loss L iou The specific formula is as follows:

[0043] ;

[0044] ;

[0045] in, L represents the number of pixels, G represents the true value map of the disguised object, δ is the indicator function, and the loss function used is L all , the specific formula is as follows:

[0046] ;

[0047] .

[0048] The beneficial effects of the present invention include:

[0049] The target detection method based on edge extraction and accurate positioning provided by the present invention constructs an EEALNet network model to set a feature extraction module, a coarse positioning module and a precise segmentation module. The feature extraction module extracts features of four different scales, and the deepest layer features are sent to the coarse positioning module. The coarse positioning module fuses the deepest layer features with other hierarchical features to obtain a coarse positioning map. The precise segmentation module further refines the coarse prediction map by alternately paying attention to the target area and the target edge to obtain the final prediction map. In the feature extraction stage, a pyramid visual converter is used to perform feature extraction. In the coarse positioning stage, the features are roughly positioned by fusing them from deep to shallow using a mirror fusion method. In the precise segmentation stage, the highly similar foreground and background are accurately separated by alternately paying attention to the target itself and the target edge. The precise positioning of the camouflaged target is gradually achieved in stages.

[0050] First, we improve ASPP and propose a new positioning module, U-ASPP (U-shaped Atrous Spatial Convolutional Pooling Pyramid). This module utilizes a U-shaped structure to fuse features extracted at different scales in ASPP from bottom to top, enhancing useful information and suppressing irrelevant or redundant information, thereby improving feature selection and effectively solving the problem of inaccurate positioning of camouflaged targets.

[0051] Second, the mirror feature fusion module uses a deep-to-shallow approach to mirror fuse features. This fusion method can effectively eliminate the huge differences between deep semantic features and shallow detail features, allowing features from two different levels to be effectively fused.

[0052] Third, the object and edge segmentation module uses dilation and erosion algorithms to expand the area around the edges of the disguised object, which helps distinguish false positive areas from false negative areas. Then, the precise positioning stage further segments the object by alternating between focusing on the object itself and its edges.

[0053] Fourth, a new positioning method is proposed in the coarse positioning stage. Features are fused from deep to shallow and the fused features are positioned, which solves the problem of deep features losing some position information. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a schematic diagram of inaccurate positioning in target detection methods in the prior art.

[0055] Figure 2 Schematic diagram of the EEALNet network model architecture of the present invention.

[0056] Figure 3 Schematic diagram of the architecture of the U-ASPP module of the present invention.

[0057] Figure 4FIG. 4 is a schematic diagram of the architecture of the MFM module of the present invention.

[0058] Figure 5 This is a schematic diagram comparing the edge extracted by the dilation algorithm of the present invention and the real edge.

[0059] Figure 6 Schematic diagram of the extracted edges of the present invention.

[0060] Figure 7 Schematic diagram of the principles of the EGRM module and FGRM module of the present invention.

[0061] Figure 8 Schematic diagram comparing the results of the EEALNet-based target detection method of the present invention with other methods. DETAILED DESCRIPTION

[0062] The following is combined with Figures 1 to 8 The present invention is described in further detail:

[0063] Example 1

[0064] See attached Figure 1 As shown, a target detection method based on edge extraction and accurate positioning includes the following steps:

[0065] S1: Constructing an EEALNet network model, wherein the EEALNet network model includes a feature extraction module, a coarse positioning module, and a precise segmentation module;

[0066] S2: The image is input into the EEALNet network model, and the feature extraction module extracts four features t1-t4 of different scales from the image;

[0067] S3: The deepest feature of the four features is sent to the coarse positioning module, which fuses the deepest feature with other hierarchical features, aligns the size of the s1 channel with the s4 channel, and reduces the number of s1 channels to 1, obtaining a coarse positioning map p1 and transmitting it to the precise segmentation module;

[0068] S4: The precise segmentation module further refines the rough prediction map by alternately paying attention to the target area and the target edge to obtain the final prediction map.

[0069] In this embodiment, the feature extraction module uses the PVTv2 network as the backbone network, and the input image is , extract four features of different scales t1-t4 in turn:

[0070] ;

[0071] Among them, I is the input image, R is a set of real numbers, H is the height, W is the width, and i is the number of layers extracted by the backbone network.

[0072] Example 2

[0073] Based on Example 1, the coarse positioning module includes an MFM module, a downsampling module, and a U-ASPP module. The specific processing process of the image by the U-ASPP module is as follows:

[0074] S31: The deepest feature t4 enters the U-ASPP module and is sent to five different branches. Through 3×3 dilated convolutions with different expansion rates, the three branches with different receptive fields extract features at multiple scales to capture objects of different sizes.

[0075] S32: The remaining two branches obtain global information through pooling and 1×1 convolution. The three branches containing dilated convolution in step S31 add the features of the previous branch after the SE attention mechanism to the features of the current branch, and then send them together to the dilated convolution and SE attention mechanism for further feature extraction;

[0076] S33: Perform channel fusion on two adjacent branches in a bottom-up order, and finally perform channel fusion on the features of the five branches to output S4.

[0077] In this embodiment, the specific processing process of the MFM module is as follows:

[0078] S34: The deep fusion feature s i+1 Further extract features through 3×3 convolution;

[0079] S35: Use Pixshuffle in super-resolution task to upsample and resize to t i Aligned, after sigmoid activation function and t i Multiply to get m1, the specific formula is as follows:

[0080] ;

[0081] in, CBR (·) represents Conv+BN+Relu, Pixshuffle (·) indicates upsampling in super-resolution;

[0082] S36: Utilize s i+1 To guide t i Focus on important features, t i Aligned by downsampling module s i+1 The size of , after sigmoid activation function and s i+1 Multiply to get m2, the specific formula is as follows:

[0083] ;

[0084] in, DownSample (·) represents a 3×3 convolution with a stride of 2;

[0085] S37: After upsampling m2 through Pixshuffle and aligning it to the size of m1, perform channel fusion with m1, and then send it to the SE attention mechanism to focus on more important features and suppress irrelevant or redundant information to obtain s i ;

[0086] ;

[0087] in, SE (·) represents the SE attention mechanism, Cat (·) represents channel fusion.

[0088] Example 3

[0089] Based on Example 1 or Example 2, the precise segmentation module includes two EGRM modules and one FGRM module. The specific processing process of the EGRM module is as follows:

[0090] See also Figure 5 The ground truth map in the dataset is subtracted from the ground truth map using the dilation algorithm to obtain the edge map P3. At the same time, the true edge map P1 in the dataset is colored to obtain P2. P2 and P3 are fused together to obtain P4. P4 shows that P3 and P2 are identical and completely overlap, proving the effectiveness and accuracy of the dilation-erosion algorithm in extracting the target edge map.

[0091] In the prior art, the expansion algorithm is used to extract the edge of the target, but the area around the edge is not paid attention to. Since there are usually false positives and false negatives around the edge, more attention should be paid to the information around the edge when extracting the target edge. Figure 6 As shown in the figure, compared with only using the dilation algorithm to extract the target edge (the green line in the figure), the dilated target is further subjected to the dilation algorithm, and then subtraction is performed to obtain the information outside the target edge (the red line in the figure). At the same time, the erosion algorithm is used on the rough prediction image to extract the information inside the target edge (the blue line in the figure). This expands the edge range and increases the information around the edge, which is conducive to distinguishing false positives and false negatives and improving the accuracy of segmentation.

[0092] The rough prediction map p i+2 Extract the edge map E through dilation and erosion algorithms i+2 , the edge graph is combined with the shallow s i+1 Multiply the shallow features s i+1It is guided and then sent to 3×3 convolution to further extract features, multiplied by the learnable parameter α, and the result is combined with f i+1 Or the results after s4 alignment size and number of channels are added and sent to CBR to further extract features and output the enhanced features f i , and finally f i Send it to the 3×3 convolution to change the number of channels to get the predicted image p i+1 , the specific formula is as follows:

[0093] ;

[0094] Where i=1 or 2, when i=2, f i+1 is s4, CBR(·) is Conv+BN+ReLU; C(·) represents channel fusion; Pixshuffle(·) represents upsampling in super-resolution; α It is a learnable parameter, initialized to 1, and Conv3 represents 3×3 convolution.

[0095] The specific processing process of the FGRM module is as follows:

[0096] Directly use the rough prediction map p i+2 As input, edges are no longer extracted, and the rest of the feature processing is the same as the EGRM module. The specific formula is as follows:

[0097] ;

[0098] Among them, i=0, CBR(·) is Conv+BN+ReLU; C(·) represents channel fusion; Pixshuffle(·) represents upsampling in super-resolution; α is a learnable parameter initialized to 1, and Conv3 represents 3×3 convolution.

[0099] The target detection method based on edge extraction and accurate positioning also includes the process of calculating the loss by setting the loss function, as follows:

[0100] will p i , i=1.2.3.4, adjust to the same size as the true value image and calculate the loss, p1 is the final output image, using binary cross entropy loss L bce Sum intersection loss L iou As a loss function. Binary cross entropy loss L bce Sum intersection loss L iou The specific formula is as follows:

[0101] ;

[0102] ;

[0103] in, L represents the number of pixels, G represents the true value map of the disguised object, δ is the indicator function, and the loss function used is L all , the specific formula is as follows:

[0104] ;

[0105] ;

[0106] To set up a verification test:

[0107] Based on four widely used COD datasets: The CHAMELEON dataset is 76 high-resolution images collected from the Internet with labels of camouflaged animals. Each image is manually annotated with object-level GT masks. The CAMO dataset contains 1,500 images with 8 categories, of which 1,000 images are used for training and 500 images are used for testing. The COD10K dataset is currently the largest dataset for camouflaged target detection, containing 5,066 images divided into 10 superclasses and 78 subclasses from multiple photography websites. The NC4K dataset consists of 4,121 images and is often used to evaluate the generalization ability of the model. A total of 4,040 images from the CAMO dataset and the COD10K dataset are divided into 1,000 images from the CAMO dataset and 3,040 images from the COD10K dataset as training sets, and all the remaining images are used as test sets.

[0108] Commonly used evaluation indicators in camouflaged target detection include structural measurement (S α ), adaptive E-measure (E φ ), F-measure, weighted F-measure ( ), mean absolute error (M). Structural measurement S α Aims to evaluate the structural information of the prediction. E-measure focuses more on pixel-level matching and image-level statistics, while F-measure focuses more on the comprehensive measurement of the accuracy and recall of the predicted image. Weighted F-measure ( ) can provide more than traditional F β More reliable evaluation results. Mean absolute error (M) computes element-wise differences.

[0109] The PyTorch framework was used to implement the model. For training, 4040 disguised target images from CAMO and COD10K were used, with the input images resized to a resolution of 384×384. The optimizer used was a stochastic gradient descent (SGD) optimizer with a momentum of 0.9 and a weight decay of 5×10−4. The model was trained with a batch size of 16, 150 epochs, and a base learning rate of 0.001. Furthermore, for testing, the input images were resized to 384×384 for model inference and then resized to their original size for evaluation.

[0110] Quantitative Analysis: Our model achieved top performance on all four datasets, significantly outperforming other methods. The F-measure metric generally improved by 1% over the next-best model on all four datasets, with a particularly strong improvement of 1.8% on the largest dataset, COD10K, demonstrating the model's high accuracy. The other four metrics also showed significant improvements over the next-best model. These significant differences further demonstrate the effectiveness and accuracy of the proposed method.

[0111] Qualitative analysis: Visual comparison with other methods Figure 8 As shown, because the coarse positioning method is changed and the disadvantage of losing position information of the deepest features is compensated, the method for locating the target is better than several other methods. For example, the positioning of the kitten in the fourth row and the fish in the first row are more accurate than the other methods. At the same time, because the edge range of interest is larger, it is also better than several other methods in edge segmentation. For example, the two targets contained in the fifth row of pictures, other methods did not separate the two targets, while the method can accurately segment each target. The method can also be seen in the second and third rows of pictures. It is better than other methods in segmentation. Therefore, the proposed network EEALNet is better than other methods in processing details, which proves the effectiveness and accuracy of the method of the present invention.

[0112] A series of ablation experiments are conducted to demonstrate the effectiveness of the three modules, namely MFM, FGRM (EGRM) and U-ASPP, and ablation experiments are also performed on multiple other modules in this article.

[0113] Table 1 Ablation experiment data of main modules

[0114]

[0115] In Table 1, B represents the base model, M represents the MFM module, S represents U-ASPP, and EF represents the EGRM and FGRM modules. Effectiveness of the MFM module: The MFM module effectively fuses two features through a mirroring method. Table 1 compares the experimental results of A and B on four datasets. B outperforms A in most metrics, demonstrating that MFM improves model performance.

[0116] Effectiveness of the EGRM (FGRM) Module: The EGRM and FGRM modules aim to further refine the rough prediction map by alternately paying attention to objects and edges, resulting in the final prediction map. Table 3 compares the experimental results of A and C on four datasets. C outperforms A in all metrics, demonstrating that the EGRM and FGRM modules can improve model performance.

[0117] Effectiveness of the U-ASPP Module: The U-ASPP module adds a U-shaped structure and SE channel attention to ASPP, aiming to roughly locate camouflaged targets through convolutions of different scales. Table 3 compares the experimental results of A and D on four datasets. D outperforms A in most metrics, demonstrating that the U-ASPP module improves model performance.

[0118] Table 2 Ablation experiments of coarse positioning method

[0119]

[0120] In Table 2, t1-t4 represent different levels of features extracted by the backbone network.

[0121] Effectiveness of the coarse localization method: As shown in Table 2, t1-t4 represent different levels of features extracted by the backbone network. Previous work primarily localizes camouflaged targets directly based on the deepest feature layer, t4. In recent years, an increasing number of studies have incorporated features from other layers for localization, resulting in improved results. In this ablation experiment, the PVT backbone network was used to extract features, followed by simple channel concatenation for feature fusion, and finally, coarse localization was performed on the fused features. As shown in Table 2, A represents discarding the t3 feature, resizing t1, t2, and t4 to the same size and number of channels through convolution, then performing channel fusion, and finally, coarse localization was performed on the fused features. The same is true for BE, except that the features used are different. In Table 2, comparing the experimental results on four datasets, F outperforms other methods in all metrics, demonstrating that our coarse localization method improves model performance.

[0122] Table 3 Ablation experiment data of EGRM (FGRM) arrangement

[0123]

[0124] In Table 3, E represents the EGRM module, and F represents the FGRM module, demonstrating the effectiveness of the EGRM (FGRM) arrangement. The EGRM and FGRM modules aim to further refine the rough prediction map by alternating attention to objects and edges, yielding the final prediction map. Comparing experimental results on four datasets, the three ADF arrangements generally outperform the three BCE arrangements, demonstrating that the proposed edge extraction method improves model performance. Furthermore, the F arrangement generally outperforms the other arrangements, demonstrating the effectiveness of the proposed EGRM (FGRM) arrangement.

[0125] Table 4 Ablation experiment data of upsampling method

[0126]

[0127] In Table 4, "Bicubic" represents bicubic interpolation upsampling, and "Pixshuffle" represents upsampling in super-resolution tasks. The effectiveness of Pixshuffle upsampling is shown in Table 4. A represents replacing all Pixshuffle upsampling methods in this paper with bicubic interpolation upsampling. Pixshuffle is an upsampling method used in super-resolution tasks that facilitates the recovery of detailed information. Comparing the experimental results of A and B on four datasets, B outperforms A in all aspects, demonstrating the accuracy of Pixshuffle upsampling compared to standard interpolation upsampling.

[0128] Table 5 Ablation experiments of the dilation algorithm times

[0129]

[0130] Effectiveness of the dilation algorithm: A represents a single dilation, extracting only the target edge. BD represent two, three, and four dilations, respectively, expanding the target edge region. The dilation algorithm aims to expand the target edge region, facilitating the identification of false positive and false negative regions around the target edge. Table 5 compares the experimental results on four datasets, showing that B outperforms A, C, and D in all aspects, demonstrating the accuracy of using the dilation algorithm twice.

[0131] This paper discovered that deep features may lose some positional information, resulting in inaccurate rough positioning of the target. Therefore, this paper uses a deep-to-shallow mirroring method to fuse features, and then roughly locates the final fused features, effectively solving the problem of inaccurate rough positioning. Also, recognizing the importance of edge information for target segmentation, the invention uses dilation and erosion algorithms to expand the target edge area, which facilitates the discrimination of false positive and false negative areas and improves segmentation accuracy. The model (EEALNet) outperforms previous methods on all benchmarks for camouflaged target detection.

[0132] In summary, the target detection method based on edge extraction and accurate positioning provided by the present invention constructs an EEALNet network model to set up a feature extraction module, a coarse positioning module and a precise segmentation module. The feature extraction module extracts four features t1-t4 of different scales, and the deepest feature is sent to the coarse positioning module. The coarse positioning module fuses the deepest feature with other hierarchical features to obtain a coarse positioning map. The precise segmentation module further refines the coarse prediction map by alternating attention to the target area and the target edge to obtain the final prediction map. In the feature extraction stage, a pyramid visual converter is used to perform feature extraction. In the coarse positioning stage, the features are roughly positioned by fusing them from deep to shallow using a mirror fusion method. In the precise segmentation stage, the highly similar foreground and background are accurately separated by alternating attention to the target itself and the target edge. The precise positioning of the camouflaged target is gradually achieved in stages.

[0133] The ASPP was improved and a new localization module, U-ASPP (U-shaped Atrous Spatial Convolutional Pooling Pyramid), was proposed. This module utilizes a U-shaped structure to fuse features extracted at different scales in the ASPP from bottom to top, enhancing useful information and suppressing irrelevant or redundant information, thereby improving feature selection and effectively addressing the issue of inaccurate localization of camouflaged targets. The mirror feature fusion module uses a deep-to-shallow mirroring method to fuse features. This fusion method effectively eliminates the significant discrepancy between deep semantic features and shallow detail features, effectively fusing features from two different levels. The object and edge segmentation module uses dilation and erosion algorithms to expand the area around the edges of camouflaged targets, facilitating the discrimination between false positive and false negative areas. The precise localization stage then further segments the target by alternating between focusing on the target itself and its edges. A new localization method is proposed in the coarse localization stage, fusing features from deep to shallow and localizing the fused features, addressing the issue of lost positional information in deep features.

Claims

1. A target detection method based on edge extraction and accurate positioning, characterized in that: The following steps are involved: S1: Constructing an EEALNet network model, wherein the EEALNet network model includes a feature extraction module, a coarse positioning module, and a precise segmentation module; S2: The image is input into the EEALNet network model, and the feature extraction module extracts four features t1-t4 of different scales from the image; S3: The deepest feature of the four features is sent to the coarse positioning module, which fuses the deepest feature with other hierarchical features, aligns the size of the s1 channel with the s4 channel, and reduces the number of s1 channels to 1, obtaining a coarse positioning map p1 and transmitting it to the precise segmentation module; S4: The precise segmentation module further refines the rough prediction map by alternately paying attention to the target area and the target edge to obtain the final prediction map; The coarse positioning module includes an MFM module, a downsampling module, and a U-ASPP module. The specific processing process of the image by the U-ASPP module is as follows: S31: The deepest feature t4 enters the U-ASPP module and is sent to five different branches. Through 3×3 dilated convolutions with different expansion rates, the three branches with different receptive fields extract features at multiple scales to capture objects of different sizes. S32: The remaining two branches obtain global information through pooling and 1×1 convolution. The three branches containing dilated convolution in step S31 add the features of the previous branch after the SE attention mechanism to the features of the current branch, and then send them together to the dilated convolution and SE attention mechanism for further feature extraction; S33: Perform channel fusion on two adjacent branches in a bottom-up order, and finally perform channel fusion on the features of the five branches to output S4; The specific processing process of the MFM module is as follows: S34: The deep fusion feature si+1 is further extracted through 3×3 convolution; S35: Use Pixshuffle in the super-resolution task for upsampling, align the size with ti, and multiply ti by the sigmoid activation function to obtain m1. The specific formula is as follows: ； Among them, CBR (·) represents Conv+BN+Relu, Pixshuffle (·) represents upsampling in super-resolution; S36: Use si+1 to guide ti to focus on important features. ti is aligned to the size of si+1 through the downsampling module, and multiplied by si+1 through the sigmoid activation function to obtain m2. The specific formula is as follows: ； Among them, DownSample(·) represents a 3×3 convolution with a stride of 2; S37: After m2 is upsampled by Pixshuffle and aligned to the size of m1, it is channel-fused with m1 and then sent to the SE attention mechanism to focus on more important features and suppress irrelevant or redundant information to obtain si; ； Among them, SE(·) represents SE attention mechanism, and Cat(·) represents channel fusion.

2. The target detection method based on edge extraction and accurate positioning according to claim 1, characterized in that: The feature extraction module uses the PVTv2 network as the backbone network, and the input image is , extract four features of different scales t1-t4 in turn: ； Among them, I is the input image, R is a set of real numbers, H is the height, W is the width, and i is the number of layers extracted by the backbone network.

3. The target detection method based on edge extraction and accurate positioning according to claim 1, characterized in that: The precise segmentation module includes two EGRM modules and one FGRM module. The specific processing process of the EGRM module is as follows: The ground truth map in the dataset is subtracted from the ground truth map using the dilation algorithm to obtain the edge map P3. At the same time, the true edge map P1 in the dataset is colored to obtain P2. P2 and P3 are fused together to obtain P4. The rough prediction map p i+2 Extract the edge map E through dilation and erosion algorithms i+2 , the edge graph is combined with the shallow s i+1 Multiply the shallow features s i+1 To guide, send it into 3×3 convolution to further extract features, multiply it with the learnable parameter α, and combine the result with f i+1 Or the results after s4 alignment size and number of channels are added and sent to CBR to further extract features and output the enhanced features f i , and finally f i Send it to the 3×3 convolution to change the number of channels to get the predicted image p i+1 , the specific formula is as follows: ； Where i=1 or 2, when i=2, f i+1 is s4, CBR(·) is Conv+BN+ReLU; C(·) represents channel fusion; Pixshuffle(·) represents upsampling in super-resolution; α It is a learnable parameter, initialized to 1, and Conv3 represents 3×3 convolution.

4. The target detection method based on edge extraction and accurate positioning according to claim 3, characterized in that: The specific processing process of the FGRM module is as follows: The rough prediction map p i+2 As input, edges are no longer extracted, and the rest of the feature processing is the same as the EGRM module. The specific formula is as follows: ； Among them, i=0, CBR(·) is Conv+BN+ReLU; C(·) represents channel fusion, Pixshuffle(·) represents upsampling in super-resolution, α is a learnable parameter initialized to 1, and Conv3 represents 3×3 convolution.

5. The target detection method based on edge extraction and accurate positioning according to claim 1, characterized in that: It also includes the process of calculating the loss by setting the loss function, as follows: will p i Adjust to the same size as the true value image and calculate the loss. p1 is the final output image, using binary cross entropy loss L bce Sum intersection loss L iou As a loss function, binary cross entropy loss L bce Sum intersection loss L iou The specific formula is as follows: ； ； in, L represents the number of pixels, G represents the true value map of the disguised object, δ is the indicator function, and the loss function used is L all , the specific formula is as follows: ； 。

Images