A small sample detection method based on effective receptive field

Abstract

The application discloses a small sample detection method based on an effective receptive field, comprising the following steps: 1) based on a query image and a support set of each category, a pre-trained visual basic model is used to extract visual features and construct a category projection space; 2) based on the query image, a pre-trained region generation network is used to generate a candidate box set in combination with the category projection space; 3) based on the candidate box, an effective receptive field support region is constructed, a category prediction result and a bounding box regression result are obtained, and a support effective receptive field region prediction model is trained; 4) in an inference stage, the effective receptive field prediction model is used to screen an optimal support region, and a final detection result is output. According to the application, a more suitable context receptive region can be adaptively screened according to a candidate target, the accuracy and stability of target classification and positioning in a small sample scene are improved, and thus the generalization ability of new category target detection and the practical application reliability are improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically a small-sample detection method based on an effective receptive field. Background Technology

[0002] Object detection is an important research area in computer vision, aiming to classify and locate objects in images. It has been widely applied in scenarios such as intelligent security, medical image analysis, remote sensing interpretation, industrial inspection, and autonomous driving. Most existing object detection methods rely on training with a large number of labeled samples, and usually achieve good results when there are sufficient samples.

[0003] However, in practical applications, many detection tasks face problems such as scarcity of new category samples, high annotation costs, and imbalanced category distribution, making traditional large-sample training methods unsuitable. Therefore, small-sample object detection has gradually gained attention, primarily using a small number of support samples to assist the model in identifying and locating new category targets.

[0004] Existing few-sample object detection methods are typically built upon existing detection frameworks. They first generate candidate bounding boxes using a candidate region generation module trained on base categories, and then combine supporting sample features for classification and regression prediction. These methods are highly dependent on base category data, and when faced with a very small number of samples and new categories of objects with significant appearance differences, they often suffer from insufficient candidate region quality and unstable detection results.

[0005] Furthermore, existing methods do not adequately utilize the contextual information surrounding the candidate bounding box, typically relying primarily on features within the box itself for discrimination. For small targets, occluded targets, targets with complex backgrounds, and targets with fine-grained differences, a mismatch between the receptive region and the actual required contextual range can easily lead to category confusion, localization bias, and decreased detection accuracy. Therefore, improving the quality of candidate regions and enhancing the effective modeling of targets and their contextual information under limited sample conditions remains a pressing technical challenge in this field. Summary of the Invention

[0006] The present invention addresses the shortcomings of the existing technology by proposing a small-sample detection method based on an effective receptive field. This method aims to improve the quality of candidate region generation for new target categories under small-sample conditions, enhance the collaborative representation ability of target features and contextual information, and thereby improve the classification accuracy, localization precision, and result stability of small-sample target detection.

[0007] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The present invention provides a small-sample detection method based on an effective receptive field, characterized by the following steps: S1. Based on the query image and the support sets corresponding to different categories, visual feature maps and category prototype vectors are obtained using a pre-trained visual base model, which are then used to generate the category projection space. ; S2. Based on the query image, a region generation network pre-trained on basic categories is used to obtain candidate feature maps of regions, and the category projection space is then applied. After linear interpolation, the results are fused with the region candidate feature map to obtain the effective receptive field candidate region feature map. ; S3. Feature map of the effective receptive field candidate region Perform foreground and bounding box prediction and filtering to generate a set of candidate boxes; S4. Based on the candidate box set, construct the effective receptive field support region of the candidate boxes and project them into the projection space of all categories. Features of candidate boxes and effective receptive field support regions are extracted and processed in the propagation layer to obtain the category prediction results and bounding box regression results of the effective receptive field support regions. S5. During the training phase, construct an effective receptive field support region prediction model, and use the prediction results of the effective receptive field support region and its category, as well as the bounding box regression results, to train the effective receptive field support region prediction model and obtain the trained effective receptive field support region prediction model. S6. During the inference phase, candidate boxes are generated. Build After identifying and processing the effective receptive field support regions, the data is input into the trained effective receptive field support region prediction model for further processing. Once the optimal effective receptive field support region is obtained, the candidate bounding boxes below the optimal effective receptive field support region are calculated using a propagation layer. The final category prediction results and bounding box regression results.

[0008] The small sample detection method based on the effective receptive field described in this invention is characterized in that step S1 includes the following steps: S11. Input the query image into the pre-trained visual base model for block encoding, feature mapping, and global feature interaction processing to obtain a visual feature map representing the semantic and spatial distribution information of the query image. ; S12. Obtain the support image set for each category from the support set corresponding to the category to be processed, and then... After feature extraction is performed on the support image set of each category input into the visual base model, the output features of the same category are then aggregated by mean to obtain the first category. The prototype vector corresponding to each category ; S13, using equation (1) to obtain the first The position of each category in the projection space Response value at Thus, the first The response values ​​of each category at each projection spatial location are used to construct the first category. The projection spaces of each category are thus obtained, thereby obtaining the projection spaces of all categories. ; (1) In equation (1), Represented as the total number of categories; For the visual feature map In position The eigenvector at that location; Spatial location index for visual feature maps; This indicates transpose.

[0009] Furthermore, step S2 includes the following steps: S21. Input the query image into a region generation network pre-trained on the basic category for feature extraction to obtain the region candidate feature map output by the last layer of the backbone network in the region generation network. ; S22. Project space according to the category. The response values ​​of each category are calculated using equation (2). The mean of the projection space corresponding to each category Based on the mean of each category, the categories are sorted, and the top K categories in the sorting results are selected. (2) S23. Perform linear interpolation mapping on the projection space corresponding to the first K categories along the category dimension to adjust its channel dimension to D dimensions, thereby obtaining the interpolated projection space. ; S24. Project the interpolation space With the candidate feature map of the region After concatenation along the channel dimension, a convolutional thinning process is performed to obtain the feature map of the effective receptive field candidate region. .

[0010] Furthermore, step S3 includes the following steps: S31. Utilizing convolutional prediction head pairs After performing convolutional feature transformation, the foreground confidence map and bounding box offset map are output respectively through the foreground classification branch and the bounding box regression branch. S32. Set predefined anchor boxes, and adjust the anchor boxes at different positions during the sliding process on the bounding box offset map to generate several initial prediction boxes; S33. Based on the foreground confidence map, sort and filter several initial prediction boxes in descending order, and filter out the initial prediction boxes corresponding to overlapping areas to obtain a set of candidate boxes.

[0011] Furthermore, step S4 includes the following steps: S41. Using any candidate box from the candidate box set Using the center point as a reference, and with the candidate box as the reference... Using the longest side as the reference scale, two candidate boxes containing the above are constructed using equation (3). Effective receptive field support area { | },in, Indicates the first A candidate box containing the above The effective receptive field support area; (3) In equation (3), for The x-coordinate of the center point for The ordinate of the center point Candidate boxes width, Candidate boxes of high; For the first One expansion ratio; S42, the candidate box and the Each effective receptive field support region is mapped to the projection space of all categories. After performing region-of-interest pooling, candidate boxes are obtained. Regional characteristics and the Effective receptive field context features of each effective receptive field support region ; S43, Using equation (4) to apply the first Effective receptive field context features With regional characteristics Processing yields the first... splicing features of effective receptive field support regions ; (4) In equation (4), Indicates splicing, Indicates the convolution operation; S44, will In the input propagation layer, the output Category prediction results and bounding box regression results .

[0012] Furthermore, step S5 includes the following steps: S51, using formula (5) The effective receptive field support region prediction model is input, and the first result is obtained using equation (5). Fit score of each effective receptive field candidate region : (5) In equation (5), To calculate the mean of the length and width dimensions, For linear layers, For activation functions; S52. Using the softmax function to... Normalized to the interval between 0 and 1, we get the first... Normalized fitness score of each effective receptive field candidate region ; S53, based on and The loss function is constructed using equation (6). This is used to update the effective receptive field support region prediction model; (6) In equation (6), Candidate boxes The index position of the true category in the category prediction result; Indicates an indicator function; Represents the cross-entropy loss function. This indicates that the first candidate box contains the aforementioned candidate box. The normalized fit score of the effective receptive field support region. This indicates that the second candidate box contains the aforementioned candidate box. The normalized fit score of the effective receptive field support region. This indicates that the first candidate box contains the aforementioned candidate box. The category prediction results of the effective receptive field support region. This indicates that the second candidate box contains the aforementioned candidate box. The effective receptive field supports the category prediction results of the region.

[0013] Furthermore, step S6 includes the following steps: S61, using candidate boxes Using the center point as a reference, the preset extended interval is generated using equation (7). Different expansion ratios : (7) In equation (7), and These represent the minimum and maximum values ​​of the expansion, respectively. Indicates the first One expansion ratio; S62, Based on M expansion ratios Calculate using equation (3) In the reasoning stage One effective receptive field support region { | },in, The first stage of reasoning One effective receptive field support area; S63, will { | Mapped to the projection space of all categories After performing region-of-interest pooling operations, the inference phase results are obtained. Effective receptive field context features { | };in, The first stage of reasoning One effective receptive field context feature; S64, will { | } and regional characteristics After splicing, the result is further processed by convolutional layers to obtain... One splicing feature { | };in, The first stage of reasoning One splicing feature; S65. Using the trained effective receptive field support region prediction model, calculate { | The fit score is calculated, and the index of the effective receptive field support region with the highest fit score is found. S66. Input the spliced ​​features of the effective receptive field support region corresponding to the index with the highest fit score into the propagation layer for processing, thereby obtaining candidate boxes. The final category prediction results and bounding box regression results.

[0014] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program supporting the processor in performing the method described therein, and the processor is configured to execute the program stored in the memory.

[0015] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program is executed by a processor to perform the steps of the method described thereon.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a category projection space and fuses it with the region candidate feature map, so that the candidate box generation process simultaneously possesses target semantic information and region localization information. This reduces the excessive reliance of existing few-shot detection methods on the basic category pre-trained candidate region generation module, improves the quality of candidate region generation for new category targets, and thus improves the accuracy and generalization ability of target detection under few-shot conditions.

[0017] 2. This invention constructs an effective receptive field support region for candidate boxes and combines it with a trained effective receptive field support region prediction model to adaptively select a context range that better matches the target. This enhances the collaborative representation of the target region and its surrounding context information, and solves the problems of fixed receptive regions and insufficient context utilization in existing methods. As a result, it improves the classification stability, localization accuracy and detection robustness of small sample target detection in complex scenarios. Attached Figure Description

[0018] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0019] In this embodiment, a small-sample detection method based on an effective receptive field is described, such as... Figure 1 As shown, it includes the following steps: In this embodiment, the query image can be a natural scene image, a medical image, a remote sensing image, or an industrial inspection image. The support set consists of a small number of labeled samples corresponding to the categories to be detected. The pre-trained visual base model can be a visual network with global feature modeling capabilities, and the region generation network can be a detection network containing a backbone network and a candidate region prediction head. During the training phase, the region generation network is first pre-trained using base category samples, and then trained using the support set and the query image together with small sample detection. During the inference phase, the query image and the support set corresponding to each category to be detected are input, and the target category and bounding box position can be output.

[0020] S1. Based on the query image and the support sets corresponding to different categories, visual feature maps and category prototype vectors are obtained using a pre-trained visual base model, which are then used to generate the category projection space. : S11. Input the query image into the pre-trained visual base model for block encoding, feature mapping, and global feature interaction processing to obtain a visual feature map representing the semantic and spatial distribution information of the query image. ; S12. Obtain the support image set for each category from the support set corresponding to the category to be processed, and then... After feature extraction is performed on the support image set of each category input into the visual base model, the output features of the same category are then aggregated by mean to obtain the first category. The prototype vector corresponding to each category .

[0021] S13, using equation (1) to obtain the first The position of each category in the projection space Response value at Thus, the first The response values ​​of each category at each projection spatial location are used to construct the first category. The projection spaces of each category are thus obtained, thereby obtaining the projection spaces of all categories. ; (1) In equation (1), Represented as the total number of categories; Visual feature map In position The eigenvector at that location; Spatial location index for visual feature maps; This indicates transpose.

[0022] S2. Based on the query image, a region generation network pre-trained on basic categories is used to obtain candidate feature maps of regions, and the category projection space is then applied. After linear interpolation, the results are fused with the region candidate feature map to obtain the effective receptive field candidate region feature map. ; S21. Input the query image into a region generation network pre-trained on the basic categories for feature extraction, and obtain the region candidate feature map output by the last layer of the backbone network in the region generation network. .

[0023] Specifically, the visual base model, as a model pre-trained in the upstream classification task, primarily represents the semantic response intensity of the category in its output category projection space, lacking the ability to explicitly model the spatial location, scale range, and boundary information of the target. Therefore, a region generation network is introduced as a candidate region providing module, and its target localization ability is learned on the basic category data, enabling it to perceive target information from the input image and predict the location and scale range of candidate regions, thereby providing region candidate feature maps for subsequent candidate box generation.

[0024] S22. Project space according to category The response values ​​of each category are calculated using equation (2). The mean of the projection space corresponding to each category Based on the mean of each category, the categories are sorted, and the top K categories in the sorting results are selected. (2) S23. Perform linear interpolation mapping on the projection space corresponding to the first K categories along the category dimension to adjust its channel dimension to D dimensions, thereby obtaining the interpolated projection space. .

[0025] S24. Project the interpolation space With region candidate feature map After concatenation along the channel dimension, a convolutional thinning process is performed to obtain the feature map of the effective receptive field candidate region. ; The values ​​for the first K categories can be set according to the category size, memory overhead, and computational complexity of the task to be detected. The channel dimension D can be set to a value that is the same as or similar to the channel dimension of the candidate feature map of the region, so as to facilitate subsequent feature fusion. By filtering, interpolating, and fusing the category projection space, we can retain the semantic information of the category with strong response on the one hand, and reduce the redundant computation and noise interference caused by the direct participation of the full category projection in the fusion on the other hand, thereby improving the discriminativeness of the candidate region features.

[0026] S3, Feature map of effective receptive field candidate regions Perform foreground and bounding box prediction and filtering to generate a set of candidate boxes; S31. Utilizing convolutional prediction head pairs After performing convolutional feature transformation, the foreground confidence map and bounding box offset map are output respectively through the foreground classification branch and the bounding box regression branch. S32. Set predefined anchor boxes and adjust the anchor boxes at different positions during the sliding process on the bounding box offset map to generate several initial prediction boxes; S33. Based on the foreground confidence map, sort and filter several initial prediction boxes in descending order, and filter out the initial prediction boxes corresponding to overlapping areas to obtain a set of candidate boxes.

[0027] S4. Based on the candidate box set, construct the effective receptive field support region of the candidate boxes and project them into the projection space of all categories. Features of candidate boxes and effective receptive field support regions are extracted and processed in the propagation layer to obtain the category prediction results and bounding box regression results of the effective receptive field support regions.

[0028] S41. Select any candidate box from the candidate box set. Using the center point as a reference, and with the candidate box... Using the longest side as the baseline, two candidate boxes are constructed using equation (3). Effective receptive field support area { | },in, Indicates the first A collection of candidate boxes The effective receptive field support area; (3) In equation (3), for The x-coordinate of the center point for The ordinate of the center point Candidate boxes width, Candidate boxes of high; For the first One expansion ratio.

[0029] Specifically, using the center point of the candidate box as the expansion benchmark helps maintain the spatial alignment between the candidate box and the context region; using the longest side of the candidate box as the benchmark scale for expansion helps to cover more complete target neighborhood information even when the aspect ratio of the candidate box varies greatly. By constructing multiple effective receptive field support regions using different expansion ratios, we can simulate the different needs of different targets for context range during the detection process, providing a candidate basis for subsequent selection of more suitable context regions.

[0030] S42, Move the candidate box and the Each effective receptive field support region is mapped to the projection space of all categories. After performing region-of-interest pooling, candidate boxes are obtained. Regional characteristics and the Effective receptive field context features of each effective receptive field support region .

[0031] S43, Using equation (4) to apply the first Effective receptive field context features With regional characteristics Processing yields the first... splicing features of effective receptive field support regions ; (4) In equation (4), Indicates splicing, This indicates a convolution operation.

[0032] S44, will In the input propagation layer, the output Category prediction results and bounding box regression results .

[0033] S5. During the training phase, construct an effective receptive field support region prediction model, and use the prediction results of the effective receptive field support region and its category, as well as the bounding box regression results, to train the effective receptive field support region prediction model and obtain the trained effective receptive field support region prediction model. S51, using formula (5) The effective receptive field support region prediction model is input, and the first result is obtained using equation (5). Fit score of each effective receptive field candidate region : (5) In equation (5), To calculate the mean of the length and width dimensions, For linear layers, This is the activation function.

[0034] S52. Using the softmax function to... Normalized to the interval between 0 and 1, we get the first... Normalized fitness score of each effective receptive field candidate region .

[0035] S53, based on and The loss function is constructed using equation (6). This is used to update the effective receptive field support region prediction model; (6) In equation (6), Candidate boxes The index position of the true category in the category prediction result; Indicates an indicator function; Represents the cross-entropy loss function. This indicates the first box containing candidate boxes. The normalized fit score of the effective receptive field support region. This indicates that the second box contains candidate boxes. The normalized fit score of the effective receptive field support region. This indicates the first box containing candidate boxes. The category prediction results of the effective receptive field support region. This indicates that the second box contains candidate boxes. The effective receptive field supports the category prediction results of the region.

[0036] In this embodiment, the training objective of the effective receptive field support region prediction model is to learn the adaptation relationship between different candidate boxes and different context ranges. For the same candidate box, the effective receptive field support regions corresponding to different expansion ratios may produce different category prediction results and bounding box regression results. By comparing their detection effects and constructing supervision signals, the effective receptive field support region prediction model can learn the mapping rule of "which context range is more conducive to the current candidate box completing the detection task", thereby providing a basis for the optimal context region selection in the inference stage.

[0037] S6. During the inference phase, candidate boxes are generated. Build After identifying and processing the effective receptive field support regions, the data is input into the trained effective receptive field support region prediction model for further processing. Once the optimal effective receptive field support region is obtained, the candidate bounding boxes below the optimal effective receptive field support region are calculated using a propagation layer. The final category prediction results and bounding box regression results.

[0038] S61, using candidate boxes Using the center point as a reference, the preset extended interval is generated using equation (7). Different expansion ratios : (7) In equation (7), and These represent the minimum and maximum values ​​of the expansion, respectively. Indicates the first One expansion ratio.

[0039] S62, Based on M expansion ratios Calculate using equation (3) In the reasoning stage One effective receptive field support region { | },in, The first stage of reasoning One effective receptive field support area; S63, will { | Mapped to the projection space of all categories After performing region-of-interest pooling operations, the inference phase results are obtained. Effective receptive field context features { | };in, The first stage of reasoning Effective receptive field context features.

[0040] Therefore, this embodiment does not uniformly process all candidate boxes with a fixed context range during the inference stage. Instead, it adaptively selects a more suitable effective receptive field region for each candidate box, and then completes the final classification and localization based on the selected region. Compared with processing methods that rely solely on local features within the candidate box or fixed-scale context regions, the above method can better balance the relationship between target ontology information and surrounding environment information, thereby improving the stability and robustness of small-sample target detection in complex scenes.

[0041] S64, will { | } and regional characteristics After splicing, the result is further processed by convolutional layers to obtain... One splicing feature { | };in, The first stage of reasoning A splicing feature.

[0042] S65. Using the trained effective receptive field support region prediction model, calculate { | The fit score is calculated, and the index of the effective receptive field support region with the highest fit score is found. S66. Input the concatenated features of the effective receptive field support region corresponding to the index with the highest fit score into the propagation layer for processing, thereby obtaining candidate boxes. The final category prediction results and bounding box regression results.

[0043] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0044] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A small-sample detection method based on an effective receptive field, characterized in that, Includes the following steps: S1. Based on the query image and the support sets corresponding to different categories, visual feature maps and category prototype vectors are obtained using a pre-trained visual base model, which are then used to generate the category projection space. ; S2. Based on the query image, a region generation network pre-trained on basic categories is used to obtain candidate feature maps of regions, and the category projection space is then applied. After linear interpolation, the results are fused with the region candidate feature map to obtain the effective receptive field candidate region feature map. ; S3. Feature map of the effective receptive field candidate region Perform foreground and bounding box prediction and filtering to generate a set of candidate boxes; S4. Based on the candidate box set, construct the effective receptive field support region of the candidate boxes and project them into the projection space of all categories. Features of candidate boxes and effective receptive field support regions are extracted and processed in the propagation layer to obtain the category prediction results and bounding box regression results of the effective receptive field support regions. S5. During the training phase, construct an effective receptive field support region prediction model, and use the prediction results of the effective receptive field support region and its category, as well as the bounding box regression results, to train the effective receptive field support region prediction model and obtain the trained effective receptive field support region prediction model. S6. During the inference phase, candidate boxes are generated. Build After identifying and processing the effective receptive field support regions, the data is input into the trained effective receptive field support region prediction model for further processing. Once the optimal effective receptive field support region is obtained, the candidate bounding boxes below the optimal effective receptive field support region are calculated using a propagation layer. The final category prediction results and bounding box regression results.

2. The small sample detection method based on the effective receptive field according to claim 1, characterized in that, Step S1 includes the following steps: S11. Input the query image into the pre-trained visual base model for block encoding, feature mapping, and global feature interaction processing to obtain a visual feature map representing the semantic and spatial distribution information of the query image. ; S12. Obtain the support image set for each category from the support set corresponding to the category to be processed, and then... After feature extraction is performed on the support image set of each category input into the visual base model, the output features of the same category are then aggregated by mean to obtain the first category. The prototype vector corresponding to each category ; S13, using equation (1) to obtain the first The position of each category in the projection space Response value at Thus, the first The response values ​​of each category at each projection spatial location are used to construct the first category. The projection spaces of each category are thus obtained, thereby obtaining the projection spaces of all categories. ; (1) In equation (1), Represented as the total number of categories; For the visual feature map In position The eigenvector at that location; Spatial location index for visual feature maps; This indicates transpose.

3. The small sample detection method based on the effective receptive field according to claim 2, characterized in that, Step S2 includes the following steps: S21. Input the query image into a region generation network pre-trained on the basic category for feature extraction to obtain the region candidate feature map output by the last layer of the backbone network in the region generation network. ; S22. Project space according to the category. The response values ​​of each category are calculated using equation (2). The mean of the projection space corresponding to each category Based on the mean of each category, the categories are sorted, and the top K categories in the sorting results are selected. (2) S23. Perform linear interpolation mapping on the projection space corresponding to the first K categories along the category dimension to adjust its channel dimension to D dimensions, thereby obtaining the interpolated projection space. ; S24. Project the interpolation space With the candidate feature map of the region After concatenation along the channel dimension, a convolutional thinning process is performed to obtain the feature map of the effective receptive field candidate region. .

4. The small sample detection method based on the effective receptive field according to claim 3, characterized in that, Step S3 includes the following steps: S31. Utilizing convolutional prediction head pairs After performing convolutional feature transformation, the foreground confidence map and bounding box offset map are output respectively through the foreground classification branch and the bounding box regression branch. S32. Set predefined anchor boxes, and adjust the anchor boxes at different positions during the sliding process on the bounding box offset map to generate several initial prediction boxes; S33. Based on the foreground confidence map, sort and filter several initial prediction boxes in descending order, and filter out the initial prediction boxes corresponding to overlapping areas to obtain a set of candidate boxes.

5. The small sample detection method based on the effective receptive field according to claim 4, characterized in that, Step S4 includes the following steps: S41. Using any candidate box from the candidate box set Using the center point as a reference, and with the candidate box as the reference... Using the longest side as the reference scale, two candidate boxes containing the above are constructed using equation (3). Effective receptive field support area { | },in, Indicates the first A candidate box containing the above The effective receptive field support area; (3) In equation (3), for The x-coordinate of the center point for The ordinate of the center point Candidate boxes width, Candidate boxes of high; For the first One expansion ratio; S42, the candidate box and the Each effective receptive field support region is mapped to a projection space of all categories. After performing region-of-interest pooling, candidate boxes are obtained. Regional characteristics and the Effective receptive field context features of each effective receptive field support region ; S43, Using equation (4) to... Effective receptive field context features With regional characteristics Processing is performed to obtain the first... splicing features of effective receptive field support regions ; (4) In equation (4), Indicates splicing, Indicates the convolution operation; S44, will In the input propagation layer, the output Category prediction results and bounding box regression results .

6. The small sample detection method based on the effective receptive field according to claim 5, characterized in that, Step S5 includes the following steps: S51, using formula (5) The effective receptive field support region prediction model is input, and the first result is obtained using equation (5). Fit score of each effective receptive field candidate region : (5) In equation (5), To calculate the mean of the length and width dimensions, For linear layers, For activation functions; S52. Using the softmax function to... Normalized to the interval between 0 and 1, we get the first... Normalized fitness score of each effective receptive field candidate region ; S53, based on and The loss function is constructed using equation (6). This is used to update the effective receptive field support region prediction model; (6) In equation (6), Candidate boxes The index position of the true category in the category prediction result; Indicates an indicator function; Represents the cross-entropy loss function. This indicates that the first candidate box contains the aforementioned candidate box. The normalized fit score of the effective receptive field support region. This indicates that the second candidate box contains the aforementioned candidate box. The normalized fit score of the effective receptive field support region. This indicates that the first candidate box contains the aforementioned candidate box. The category prediction results of the effective receptive field support region. This indicates that the second candidate box contains the aforementioned candidate box. The category prediction results of the effective receptive field support region.

7. The small sample detection method based on the effective receptive field according to claim 6, characterized in that, Step S6 includes the following steps: S61, using candidate boxes Using the center point as a reference, the preset extended interval is generated using equation (7). Different expansion ratios : (7) In equation (7), and These represent the minimum and maximum values ​​of the expansion, respectively. Indicates the first One expansion ratio; S62, Based on M expansion ratios Calculate using equation (3) In the reasoning stage One effective receptive field support region { | },in, The first stage of reasoning One effective receptive field support area; S63, will { | Mapped to the projection space of all categories After performing region-of-interest pooling operations, the inference phase results are obtained. Effective receptive field context features { | };in, The first stage of reasoning One effective receptive field context feature; S64, will { | } and regional characteristics After splicing, the result is further processed by convolutional layers to obtain... One splicing feature { | };in, The first stage of reasoning One splicing feature; S65. Using the trained effective receptive field support region prediction model, calculate { | The fit score is calculated, and the index of the effective receptive field support region with the highest fit score is found. S66. Input the spliced ​​features of the effective receptive field support region corresponding to the index with the highest fit score into the propagation layer for processing, thereby obtaining candidate boxes. The final category prediction results and bounding box regression results.

8. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the method of any one of claims 1-7, the processor being configured to execute the program stored in the memory.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by the processor to perform the steps of the method according to any one of claims 1-7.

Images