Dead chicken detection method based on multi-scale reconstruction layer aggregation and feature compensation progressive fusion

Abstract

The application is suitable for the technical field of target detection, and provides a dead chicken detection method based on multi-scale reconstruction layer aggregation and feature compensation progressive fusion, including: acquiring a target detection image, the target detection image including dead chickens and live chickens, inputting the target detection image into a preset YOLO model, outputting a dead chicken detection result, the preset YOLO model including a backbone feature extraction network, a feature fusion module and a detection output head, the backbone feature extraction network being a preset multi-scale reconstruction efficient layer aggregation MSR-ELAN network, and the feature fusion module being a preset feature compensation progressive fusion pyramid FCPFN network. Thus, by segmenting and recombining the extracted features, using a multi-stage feature interaction and integration strategy, the overall perception ability of the target is optimized, thereby effectively improving the detection performance of the dead chicken target.

Description

Technical Field

[0001] This application belongs to the field of target detection technology, and in particular relates to a dead chicken detection method based on multi-scale reconstruction layer aggregation and feature compensation progressive fusion. Background Technology

[0002] Currently, my country's broiler industry has begun to transform and upgrade towards large-scale, intelligent, and green operations, but it has not yet achieved unmanned farming. Most chicken farms still rely on manual inspections to find and remove sick chickens. This method has drawbacks such as high labor intensity, strong subjectivity, and the potential for sick chickens to remain due to visual fatigue among farm workers. As is well known, dead chickens left untreated for extended periods can generate large amounts of bacteria, leading to large-scale influenza outbreaks.

[0003] In related technologies, current target detection algorithms applied to dead chicken detection suffer from insufficient dead chicken target features and feature loss due to occlusion, resulting in insufficient dead chicken detection accuracy. Summary of the Invention

[0004] This application provides a dead-chicken detection method based on multi-scale reconstruction layer aggregation and feature compensation progressive fusion. This method addresses the problem that current target detection algorithms used for dead-chicken detection suffer from insufficient dead-chicken target features and feature loss due to occlusion, resulting in insufficient dead-chicken detection accuracy.

[0005] In a first aspect, embodiments of this application provide a dead chicken detection method based on multi-scale reconstruction layer aggregation and feature compensation progressive fusion, comprising: acquiring a target detection image, wherein the target detection image includes dead chickens and live chickens; inputting the target detection image into a preset YOLO model, and outputting dead chicken detection results, wherein the preset YOLO model includes a backbone feature extraction network, a feature fusion module, and a detection output head, wherein the backbone feature extraction network is a preset multi-scale reconstruction high-efficiency layer aggregation MSR-ELAN network, and the feature fusion module is a preset feature compensation progressive fusion pyramid FCPPN network.

[0006] In one possible implementation of the first aspect, the above-mentioned inputting the target detection image into a preset YOLO model and outputting dead chicken detection results includes:

[0007] The target detection image is input into a preset multi-scale reconstruction high-efficiency layer aggregation MSR-ELAN network, which outputs feature maps of multiple sizes.

[0008] Multiple feature maps of different sizes are input into a pre-defined feature-compensated progressive fusion pyramid FPFN network, which outputs multiple fused feature maps.

[0009] Multiple fused feature maps are input into the detection output head, and the dead chicken detection result is output.

[0010] Optionally, in another possible implementation of the first aspect, the preset multi-scale reconstruction efficient layer aggregation MSR-ELAN network includes four layers, each layer including a multi-scale reconstruction extraction block. The target detection image is input into the preset multi-scale reconstruction efficient layer aggregation MSR-ELAN network, and the output feature maps of multiple sizes are included:

[0011] The target detection image is processed through standard convolution to generate a basic feature map;

[0012] The basic feature map is input into the first-level multi-scale reconstruction extraction block, and the first feature map is output.

[0013] The first feature map is processed by standard convolution to output the first convolutional feature map;

[0014] After adjusting the number of channels of the first convolutional feature map through standard convolution, the number of channels of the first convolutional feature map is adjusted to C1;

[0015] The first feature map is input into the second-level multi-scale reconstruction extraction block, and the second feature map is output.

[0016] The second feature map is processed by standard convolution to output the second convolutional feature map;

[0017] After adjusting the number of channels of the second convolutional feature map through standard convolution, the number of channels of the second convolutional feature map is adjusted to C1+C2, and the second convolutional feature map is segmented according to the number of channels C1+C2 to generate the second convolutional feature map C1 and the second convolutional feature map C2.

[0018] The second feature map is input into the multi-scale reconstruction extraction block of the third level, and the third feature map is output.

[0019] The third feature map is processed by standard convolution to output the third convolutional feature map;

[0020] After adjusting the number of channels of the third convolutional feature map through standard convolution, the number of channels of the third convolutional feature map is adjusted to C1+C2+C3. Then, the third convolutional feature map is segmented according to the number of channels C1+C2+C3 to generate the third convolutional feature map C1, the third convolutional feature map C2 and the third convolutional feature map C3.

[0021] The third feature map is input into the fourth-level multi-scale reconstruction extraction block, and the fourth feature map is output.

[0022] The fourth feature map is processed by standard convolution to output the fourth convolutional feature map;

[0023] After adjusting the number of channels of the fourth convolutional feature map through standard convolution, the number of channels of the fourth convolutional feature map is adjusted to C1+C2+C3+C4. Then, the fourth convolutional feature map is segmented according to the number of channels C1+C2+C3+C4 to generate the fourth convolutional feature map C1, the fourth convolutional feature map C2, the fourth convolutional feature map C3 and the fourth convolutional feature map C4.

[0024] In the first level, the second convolutional feature map C1, the third convolutional feature map C1, and the fourth convolutional feature map C1 are upsampled and then combined with the first convolutional feature map to generate a feature map of the first size.

[0025] In the second level, the third convolutional feature map C2 and the fourth convolutional feature map C2 are upsampled and then combined with the second convolutional feature map C2 to generate a feature map of the second size.

[0026] In the third level, the fourth convolutional feature map C3 is upsampled and then combined with the third convolutional feature map C3 to generate a feature map of the third size.

[0027] In the fourth level, the fourth convolutional feature map is determined as the feature map of the fourth size.

[0028] Optionally, in another possible implementation of the first aspect, the above-described input of the basic feature map to the multi-scale reconstruction extraction block of the first level, and output of the first feature map, includes:

[0029] After convolution processing, the basic feature map is divided into a first basic feature sub-map and a second basic feature sub-map.

[0030] The first basic feature sub-image is convolved to generate the first basic convolutional feature sub-image.

[0031] The second basic feature sub-image is subjected to multi-scale reconstruction convolution processing and repeated multiple times to generate the second basic convolution feature sub-image.

[0032] The first basic convolutional feature sub-image and the second basic convolutional feature sub-image are concatenated and convolved sequentially to generate the first feature map.

[0033] Optionally, in another possible implementation of the first aspect, the above-described multi-scale reconstruction convolutional processing of the second basic feature sub-map includes:

[0034] The second basic feature sub-map is sequentially processed by convolution, batch normalization, and non-linear activation by the ReLU6 activation function to generate an extended activation feature map.

[0035] The extended activation feature map is processed by multiple structural reconstruction convolutions (SCConv) simultaneously to generate multiple reconstructed feature maps. The scale of the convolution process is different in each SCConv process.

[0036] Multiple reconstructed feature maps are batch normalized, non-linearly activated by the ReLU6 activation function, and then fused. Finally, channel shuffling and convolution are performed.

[0037] Optionally, in another possible implementation of the first aspect, the above-described extended activation feature map is processed by multiple structural reconstruction convolutional SCConv operations simultaneously to generate multiple reconstructed feature maps, including:

[0038] Each time, the expanded activation feature map is processed sequentially by spatial reconstruction SRU, channel reconstruction CRU, and convolution to generate a reconstructed feature map.

[0039] Optionally, in another possible implementation of the first aspect, the above-mentioned input of feature maps of multiple sizes into a preset feature-compensated progressive fusion pyramid (FCPFN) network outputs multiple fused feature maps, including:

[0040] In order of size, feature maps of every two adjacent pairs are subjected to feature compensation processing and then standard convolution processing to generate compensated feature maps, including:

[0041] Feature compensation processing is performed based on the feature map of the first size and the feature map of the second size, followed by standard convolution processing to generate the first compensated feature map.

[0042] Feature compensation is performed based on the feature maps of the second and third dimensions, followed by standard convolution to generate a second compensated feature map.

[0043] Feature compensation is performed based on the feature maps of the third and fourth dimensions, followed by standard convolution to generate the third compensated feature map.

[0044] Perform standard convolution on the feature map of the fourth size to generate the fourth compensated feature map;

[0045] After upsampling the second compensation feature map, it is adaptively fused with the first compensation feature map and then subjected to standard convolution to generate a fused feature map A1.

[0046] After downsampling the first compensation feature map, the second compensation feature map is adaptively fused and then subjected to standard convolution to generate the fused feature map B1.

[0047] After downsampling the fused feature map B1 and fused feature map A1, adaptive fusion is performed with the third compensation feature map, followed by convolution to generate fused feature map C1.

[0048] After upsampling the fused feature map B1 and the third compensation feature map, they are adaptively fused with the fused feature map A1 and then convolved to generate the fused feature map A2.

[0049] After downsampling and upsampling of the fused feature map A1 and the third compensation feature map respectively, they are adaptively fused with the fused feature map B1 and then convolved to generate the fused feature map B2.

[0050] After upsampling the fused feature map B2, fused feature map C1, and the fourth compensation feature map, adaptive fusion is performed with the fused feature map A2 to generate a first-size fused feature map.

[0051] After downsampling the fused feature map A2, upsampling the fused feature map C1 and the fourth compensation feature map, and adaptively fusing them with the fused feature map B2, a second-size fused feature map is generated.

[0052] After downsampling the fused feature map A2 and fused feature map B2, and upsampling the fourth compensation feature map, adaptive fusion is performed with the fused feature map C1 to generate a third-size fused feature map.

[0053] After downsampling the fused feature maps A2, B2, and C1, adaptive fusion is performed with the fourth compensation feature map to generate a fourth-size fused feature map.

[0054] Optionally, in another possible implementation of the first aspect, the above-described method of performing feature compensation processing on every two adjacent feature maps in order of size, followed by standard convolution processing, to generate compensated feature maps includes:

[0055] After performing convolution and upsampling on the first feature map, the difference between it and the second feature map after convolution is calculated to obtain the lost key information.

[0056] The missing key information is added to the second feature map after convolution to generate a compensation feature map.

[0057] Optionally, in another possible implementation of the first aspect, the above adaptive fusion includes:

[0058] Assign corresponding weights to the multiple feature maps for adaptive fusion;

[0059] Based on the weights corresponding to each feature map, the multiple feature maps that are adaptively fused are weighted and fused.

[0060] Beneficial Effects: In the technical solution of this application, target detection images are first acquired, including both dead and live chickens. These images are then input into a preset YOLO model, which outputs the detection results for dead chickens. The preset YOLO model includes a backbone feature extraction network, a feature fusion module, and a detection output head. The backbone feature extraction network is a preset multi-scale reconstruction high-efficiency layer aggregation MSR-ELAN network, and the feature fusion module is a preset feature-compensated progressive fusion pyramid FCPPN network. Therefore, by segmenting and recombining the extracted features and utilizing a multi-stage feature interaction and integration strategy, the overall perception capability of the target is optimized, thereby effectively improving the detection performance of dead chicken targets. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is a flowchart illustrating a dead chicken detection method based on multi-scale reconstruction layer aggregation and feature compensation progressive fusion provided in an embodiment of this application.

[0063] Figure 2 This is a schematic diagram of the structure of an MSR-ELAN network provided in one embodiment of this application;

[0064] Figure 3 This is a schematic diagram of the structure of a multi-scale reconstruction extraction block provided in an embodiment of this application;

[0065] Figure 4 This is a schematic diagram of the structure of a multi-scale reconstruction convolution provided in an embodiment of this application;

[0066] Figure 5 This is a schematic diagram of the structure of a structure reconstruction convolution provided in an embodiment of this application;

[0067] Figure 6 This is a schematic diagram of the structure of a preset feature-compensated progressive fusion pyramid network provided in an embodiment of this application;

[0068] Figure 7 This is a schematic diagram of the downsampling structure provided in an embodiment of this application;

[0069] Figure 8 This is a schematic diagram of the feature compensation module provided in an embodiment of this application;

[0070] Figure 9 This is a schematic diagram of the target detection visualization results provided in an embodiment of this application by MSR-FCPN. Detailed Implementation

[0071] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0072] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0073] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0074] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0075] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0076] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0077] The following description, with reference to the accompanying drawings, details the dead-chicken detection method based on multi-scale reconstruction layer aggregation and feature compensation progressive fusion provided in this application.

[0078] Figure 1 The diagram shows a flowchart of a dead chicken detection method based on multi-scale reconstruction layer aggregation and feature compensation progressive fusion provided in an embodiment of this application.

[0079] like Figure 1 As shown, the dead chicken detection method based on multi-scale reconstruction layer aggregation and feature compensation progressive fusion includes the following steps:

[0080] S1, Obtain the target detection image, which includes dead chickens and live chickens;

[0081] S2, input the target detection image into the preset YOLO model and output the dead chicken detection result. The preset YOLO model includes a backbone feature extraction network, a feature fusion module and a detection output head. The backbone feature extraction network is a preset multi-scale reconstruction high-efficiency layer aggregation MSR-ELAN network, and the feature fusion module is a preset feature compensation progressive fusion pyramid FCPPN network.

[0082] This application proposes the MSR-FCPN framework for detecting small and occluded targets in dead chickens in multi-layer caged broiler environments. It still adopts the three-segment architecture of the YOLO model, namely the Backbone feature extraction network, the Neck feature fusion module, and the Head detection output head. The improvements and optimizations are as follows: (1) In the Backbone part, a new multi-scale reconstruction efficient layer aggregation network (MSR-ELAN) is proposed as the core operation unit. This network uses the multi-scale reconstruction extraction block (MSEB) to perform feature extraction operations and introduces feature segmentation and reconstruction mechanisms to comprehensively obtain feature information of different scales in the image. MSEB uses multi-scale reconstruction convolution (MSRC) as the key component and uses convolution kernels of different sizes to process the input feature map in parallel. Among them, the smaller convolution kernel focuses on capturing the subtle local features of the target, while the larger convolution kernel is responsible for obtaining a broader spatial context information. The two work together to ensure the comprehensiveness of feature extraction. (2) In the Neck structure, a feature compensation progressive fusion pyramid network (FCPFN) is proposed to further process the multi-scale feature map extracted by MSEB. The feature compensation module within FCPFN can leverage the differences between feature maps at different scales to compensate for potentially lost critical dead-end feature information. The compensated feature information is not directly integrated, but rather fused gradually, starting from adjacent layers, to progressively reduce the semantic gap between features at different levels, effectively preventing the loss of feature information during cross-layer transmission and interaction.

[0083] Furthermore, in this embodiment of the application, S2 includes:

[0084] S21, Input the target detection image into the preset multi-scale reconstruction high-efficiency layer aggregation MSR-ELAN network, and output feature maps of multiple sizes;

[0085] It should be noted that in multi-layered cage farming environments, the occlusion of live chickens can partially obscure the visible portion of dead chickens, reducing the scale of the dead chicken in the image. Furthermore, the dynamic activity of live chickens causes the occlusion relationship to change over time, further increasing the uncertainty of the target scale. To address this complexity, the backbone feature extraction network needs strong multi-scale feature perception capabilities to comprehensively capture feature information at different scales. To enhance the feature extraction capabilities of the base network, this application designs a Multi-Scale Reconstruction Efficient Layer Aggregation Network (MSR-ELAN). The output of each layer is combined with the input of the next layer, significantly improving the MSR-FCPN model's ability to perceive features at different scales while maintaining computational efficiency and retaining more key target information.

[0086] Furthermore, in the embodiments of this application, according to Figure 2 As shown, the preset multi-scale reconstruction high-efficiency layer aggregation MSR-ELAN network includes four layers, each layer including a multi-scale reconstruction extraction block; the above S21 includes:

[0087] The target detection image is processed through standard convolution to generate a basic feature map;

[0088] The basic feature map is input into the first-level multi-scale reconstruction extraction block, and the first feature map is output.

[0089] The first feature map is processed by standard convolution to output the first convolutional feature map;

[0090] After adjusting the number of channels of the first convolutional feature map through standard convolution, the number of channels of the first convolutional feature map is adjusted to C1;

[0091] The first feature map is input into the second-level multi-scale reconstruction extraction block, and the second feature map is output.

[0092] The second feature map is processed by standard convolution to output the second convolutional feature map;

[0093] After adjusting the number of channels of the second convolutional feature map through standard convolution, the number of channels of the second convolutional feature map is adjusted to C1+C2, and the second convolutional feature map is segmented according to the number of channels C1+C2 to generate the second convolutional feature map C1 and the second convolutional feature map C2.

[0094] The second feature map is input into the multi-scale reconstruction extraction block of the third level, and the third feature map is output.

[0095] The third feature map is processed by standard convolution to output the third convolutional feature map;

[0096] After adjusting the number of channels of the third convolutional feature map through standard convolution, the number of channels of the third convolutional feature map is adjusted to C1+C2+C3. Then, the third convolutional feature map is segmented according to the number of channels C1+C2+C3 to generate the third convolutional feature map C1, the third convolutional feature map C2 and the third convolutional feature map C3.

[0097] The third feature map is input into the fourth-level multi-scale reconstruction extraction block, and the fourth feature map is output.

[0098] The fourth feature map is processed by standard convolution to output the fourth convolutional feature map;

[0099] After adjusting the number of channels of the fourth convolutional feature map through standard convolution, the number of channels of the fourth convolutional feature map is adjusted to C1+C2+C3+C4. Then, the fourth convolutional feature map is segmented according to the number of channels C1+C2+C3+C4 to generate the fourth convolutional feature map C1, the fourth convolutional feature map C2, the fourth convolutional feature map C3 and the fourth convolutional feature map C4.

[0100] In the first level, the second convolutional feature map C1, the third convolutional feature map C1, and the fourth convolutional feature map C1 are upsampled and then combined with the first convolutional feature map to generate a feature map of the first size.

[0101] In the second level, the third convolutional feature map C2 and the fourth convolutional feature map C2 are upsampled and then combined with the second convolutional feature map C2 to generate a feature map of the second size.

[0102] In the third level, the fourth convolutional feature map C3 is upsampled and then combined with the third convolutional feature map C3 to generate a feature map of the third size.

[0103] In the fourth level, the fourth convolutional feature map is determined as the feature map of the fourth size.

[0104] It should be noted that the number of channels C1, C2, C3, and C4 are preset number of channels, and this application does not limit them.

[0105] It should be noted that the above upsampling process Figure 2 As shown, in the MSR-ELAN network, it can be implemented as an upsampling convolutional block. The upsampling process includes: upsampling N×, depthwise convolution 3×3, batch normalization, ReLU activation function, and convolutional layer.

[0106] Furthermore, in the embodiments of this application, according to Figure 3As shown, the above-mentioned input of the basic feature map into the first-level multi-scale reconstruction extraction block and output of the first feature map includes:

[0107] After convolution processing, the basic feature map is divided into a first basic feature sub-map and a second basic feature sub-map.

[0108] The first basic feature sub-image is convolved to generate the first basic convolutional feature sub-image.

[0109] The second basic feature sub-image is subjected to multi-scale reconstruction convolution processing and repeated multiple times to generate the second basic convolution feature sub-image.

[0110] The first basic convolutional feature sub-image and the second basic convolutional feature sub-image are concatenated and convolved sequentially to generate the first feature map.

[0111] In this embodiment, the implementation process of MSEB is as follows: The basic feature map is divided into two parts. One part is directly passed through ordinary convolution operations, while the other part is extracted using multiple MSRCs for deep multi-scale feature extraction. Finally, the two feature parts are concatenated and fused through convolution to form the final feature map. The feature map not only retains the dead chicken information in the original feature map, but also includes feature information captured from different scales after deep processing by MRCB, enabling the model to more comprehensively understand the image content.

[0112] It should be noted that the above-mentioned multi-scale reconstruction convolution process of the second basic feature sub-image is repeated multiple times, and the number of times it is executed is not limited in this application.

[0113] It should also be noted that the processes described above—inputting the first feature map into the second-level multi-scale reconstruction extraction block and outputting the second feature map, inputting the second feature map into the third-level multi-scale reconstruction extraction block and outputting the third feature map, and inputting the third feature map into the fourth-level multi-scale reconstruction extraction block and outputting the fourth feature map—are consistent with the processes described in the embodiments of this application, and will not be repeated here.

[0114] Furthermore, in the embodiments of this application, according to Figure 4 As shown, the above-mentioned multi-scale reconstruction convolutional processing of the second basic feature sub-image specifically includes:

[0115] The second basic feature sub-map is sequentially processed by convolution, batch normalization, and non-linear activation by the ReLU6 activation function to generate an extended activation feature map.

[0116] The extended activation feature map is processed by multiple structural reconstruction convolutions (SCConv) simultaneously to generate multiple reconstructed feature maps. The scale of the convolution process is different in each SCConv process.

[0117] Multiple reconstructed feature maps are batch normalized, non-linearly activated by the ReLU6 activation function, and then fused. Finally, channel shuffling and convolution are performed.

[0118] In this embodiment, the multi-scale reconstruction convolutional processing is implemented as follows: First, pointwise convolutional layers are used to expand the number of channels, enhancing the expressive power of features. Next, batch normalization (BN) layers are used to normalize the features, and ReLU6 activation layers (R6) are used for non-linear activation to reduce internal covariate shifts. Subsequently, SCConv layers with different kernel sizes are used to extract features from the feature map at multiple scales. The refined features are then integrated together again through BN and ReLU6 layers. After feature extraction, a channel shuffle operation is used to merge the relationships between channels, ensuring more complete feature fusion. Then, pointwise convolutions are used to compress the number of channels back to the original number of channels. This process not only reduces computational complexity but also preserves important feature information.

[0119] Furthermore, in the embodiments of this application, according to Figure 5 As shown, the above process involves performing multiple structural reconstruction convolutional SCConv operations on the expanded activation feature map simultaneously, generating multiple reconstructed feature maps, including:

[0120] Each time, the expanded activation feature map is processed sequentially by spatial reconstruction SRU, channel reconstruction CRU, and convolution to generate a reconstructed feature map.

[0121] It should be noted that SCConv is an innovative convolutional neural network compression method that can improve the feature extraction performance of the model while reducing spatial and channel redundancy in convolutional layers. Figure 5 As shown, SCConv consists of two parts in sequence: Spatial Reconstruction Unit (SRU) and Channel Reconstruction Unit (CRU). The SRU separates redundant features based on weights and reconstructs them to suppress redundancy in the spatial dimension and enhance representation capabilities. The CRU utilizes splitting, transforming, and fusion strategies to reduce redundancy in the channel dimension, as well as computational cost and storage requirements. SCConv fully leverages the redundancy of intermediate feature maps, reducing the number of parameters and computational load without compromising performance.

[0122] S22, input feature maps of multiple sizes into the preset feature-compensated progressive fusion pyramid FPFN network, and output multiple fused feature maps;

[0123] In multi-layer caged broiler environments, the detection of small dead chicken targets and occluded dead chicken targets faces the challenge of selecting the appropriate input feature map size. An excessively large input feature map significantly increases computation time, reducing model efficiency, while an excessively small feature map can lead to the loss of key target information. Furthermore, the semantic gap between features at non-adjacent levels is greater than that between adjacent levels, especially between bottom and top features, resulting in poor feature fusion performance at non-adjacent levels. To address this, this paper proposes a Feature Compensation Progressive Fusion Pyramid Network (FCPFN), which comprises feature compensation and progressive fusion, effectively improving detection performance. Feature compensation enhances the feature representation of dead chicken targets; progressive fusion reduces semantic gaps by gradually fusing features from different levels. During progressive fusion, an adaptive spatial fusion operation is used to suppress information contradictions between features at different levels, thereby effectively improving the model's ability to detect dead chicken targets.

[0124] In the embodiments of this application, such as Figure 6 As shown, the above S22 includes:

[0125] S2201, in order of size, performs feature compensation processing on every two adjacent feature maps, and then performs standard convolution processing to generate compensated feature maps, including:

[0126] (1) Perform feature compensation processing based on the feature map of the first size and the feature map of the second size, and perform standard convolution processing to generate the first compensated feature map;

[0127] (2) Perform feature compensation processing based on the feature map of the second size and the feature map of the third size, and perform standard convolution processing to generate the second compensated feature map;

[0128] (3) Perform feature compensation processing based on the feature map of the third size and the feature map of the fourth size, and perform standard convolution processing to generate the third compensated feature map;

[0129] (4) Perform standard convolution on the feature map of the fourth size to generate the fourth compensated feature map;

[0130] S2202, after upsampling the second compensation feature map, it is adaptively fused with the first compensation feature map and then subjected to standard convolution to generate the fused feature map A1;

[0131] S2203, after downsampling the first compensation feature map, adaptively fuse the second compensation feature map and perform standard convolution to generate the fused feature map B1;

[0132] S2204, after downsampling the fused feature map B1 and the fused feature map A1, adaptively fuse them with the third compensation feature map and perform convolution processing to generate the fused feature map C1;

[0133] S2205, after upsampling the fused feature map B1 and the third compensation feature map, adaptively fuse them with the fused feature map A1 and perform convolution processing to generate the fused feature map A2;

[0134] S226, After downsampling and upsampling of the fused feature map A1 and the third compensation feature map respectively, adaptive fusion is performed with the fused feature map B1, and convolution is performed to generate the fused feature map B2;

[0135] S2207, after upsampling the fused feature map B2, fused feature map C1, and the fourth compensation feature map, adaptive fusion is performed with the fused feature map A2 to generate a first-size fused feature map;

[0136] S2208, after downsampling the fused feature map A2, upsampling the fused feature map C1 and the fourth compensation feature map, and adaptively fusing them with the fused feature map B2, a second-size fused feature map is generated;

[0137] S2209, after downsampling the fused feature map A2 and the fused feature map B2, upsampling the fourth compensation feature map, and adaptively fusing it with the fused feature map C1, a third-size fused feature map is generated.

[0138] S2210, after downsampling the fused feature maps A2, B2, and C1, adaptively fuse them with the fourth compensation feature map to generate a fourth-size fused feature map.

[0139] It should be noted that, according to Figure 7 As shown, downsampling includes: N× downsampling, 3×3 depthwise convolution, batch normalization layer, ReLU activation function, and convolutional layer.

[0140] Furthermore, in the embodiments of this application, according to Figure 8 As shown, step S2201 above includes:

[0141] After performing convolution and upsampling on the first feature map, the difference between it and the second feature map after convolution is calculated to obtain the lost key information.

[0142] The missing key information is added to the second feature map after convolution to generate a compensation feature map.

[0143] Optionally, in one embodiment, the feature compensation module of FCPPN fully utilizes the size and semantic differences between deep and shallow feature maps to enhance the features of the dead chicken target. The implementation process of the feature compensation module is as follows: First, the input image is processed by a Backbone network to extract multi-scale feature maps. Taking the feature maps of the last two layers as an example, the penultimate layer is denoted as A1, which has a small size but rich semantic information; the penultimate layer is denoted as A2, which has a large size and contains rich target detail information. A1 is restored to the same size as A2 through point convolution and upsampling modules, while A2 is processed by point convolution. Then, the difference between the A1 feature map and the A2 feature map of the same size is calculated to obtain the lost dead chicken key target information F. lost .

[0144] F lost =UpSample(Conv 1×1 (A1))-Conv 1×1 (A2)

[0145] Key information about the missing dead chicken F lost Add the result to A2 to output the compensated feature map, thereby enhancing A2's feature representation of the dead chicken target.

[0146] Output = F lost +A2

[0147] By compensating for the lost dead chicken target information, the details of the shallow feature map are highlighted, which helps the model better understand the appearance and shape of the target, significantly enhances the feature representation of the dead chicken target, and effectively prevents the loss of important feature information in the subsequent feature fusion process.

[0148] Furthermore, in this embodiment, the adaptive fusion in S2202-S2210 includes:

[0149] Assign corresponding weights to the multiple feature maps for adaptive fusion;

[0150] Based on the weights corresponding to each feature map, the multiple feature maps that are adaptively fused are weighted and fused.

[0151] In one embodiment, with the assistance of the feature compensation module, FCPPN implements a progressive fusion strategy for feature maps to achieve deep integration and optimization. Taking four feature maps as an example, the strategy consists of three stages. First, the low-level features A3 and A4 are initially fused; low-level features typically have rich detailed information and relatively accurate spatial location information, but relatively little semantic information. Through initial fusion, the advantages of low-level features can be fully utilized, laying the foundation for subsequent feature fusion. Then, the deep feature A2 is gradually incorporated; deep features have richer semantic information, which can help the model better understand the image content. The incorporation of A2 further enriches the semantic information of the features. Finally, the top-level, most abstract feature A1 is fused; the A1 feature has high semantic abstraction and can provide global semantic information. The incorporation of A1 can further enhance the expressive power of the features, enabling the model to understand the image content more comprehensively. This strategy can not only effectively narrow the semantic gap between features at different levels, but also fully utilize multi-scale feature information, thereby comprehensively improving the detection performance of the model. To further improve the feature fusion effect, this paper introduces an adaptive spatial fusion (ASFF) mechanism, which significantly enhances the importance of key-level features while effectively suppressing contradictory information between different objects. This mechanism assigns different spatial weights to features at different levels according to Equations 1 and 2, and then fuses bottom-level, high-level, and top-level features sequentially.

[0152]

[0153] in, This represents the feature vector (n∈1,2,3,4) from position (i,j) in the nth to kth layers. This represents the feature vector obtained through adaptive spatial fusion. and These represent the spatial weights of the four layer features of the k-th layer.

[0154] S23, input multiple fused feature maps to the detection output head, and output the dead chicken detection result.

[0155] This application provides a dead chicken detection method based on multi-scale reconstruction layer aggregation and feature compensation progressive fusion. First, a target detection image is acquired, which includes both dead and live chickens. The target detection image is then input into a pre-defined YOLO model, which outputs dead chicken detection results. The pre-defined YOLO model includes a backbone feature extraction network, a feature fusion module, and a detection output head. The backbone feature extraction network is a pre-defined multi-scale reconstruction high-efficiency layer aggregation MSR-ELAN network, and the feature fusion module is a pre-defined feature compensation progressive fusion pyramid FCPPN network. Therefore, by segmenting and recombining the extracted features and utilizing a multi-stage feature interaction and integration strategy, the overall perception capability of the target is optimized, thereby effectively improving the detection performance of dead chicken targets.

[0156] The above embodiments demonstrate the overall process of the dead chicken detection method proposed in this application, which is based on multi-scale reconstruction layer aggregation and feature compensation progressive fusion. To comprehensively evaluate the effectiveness of the proposed MSR-FCPN model, considering the uncertainty of the dead chicken's location in the image, a Mosaic data augmentation method is employed to scale, flip, and stitch the images to enrich the training data. The dataset contains 1,348 valid images, which are randomly divided into training, validation, and test sets in an 8:1:1 ratio.

[0157] Comparative experiment:

[0158] In selecting comparison methods, this application chose cutting-edge methods in the field of object detection as references, covering the advanced YOLO series, including YOLOv5, PP-YOLOE+, YOLOv6-3.0, Gold-YOLO, YOLOv8, YOLO-MS, YOLOv9, YOLOv10, and YOLOv11. To ensure experimental fairness, all comparison methods used their default parameter configurations, and the influence of pre-training and self-distillation strategies was excluded. Furthermore, to reduce the interference of input image size on the results, the experimental input resolution was uniformly standardized to 640×640 pixels, a common standard in the field.

[0159] During the experiments, this application not only focused on the detection accuracy of the model, but also comprehensively considered the model's computational complexity and number of parameters. The model's performance was evaluated using four metrics: accuracy (P), recall (R), mean precision (mAP@0.5), and mAP0.5–0.95. Simultaneously, the model's GFLOPs (billions of floating-point operations per second) were recorded to reflect computational complexity, and the Params value (number of model parameters) was used to assess the model's storage requirements and potential overfitting risk.

[0160] Table 1

[0161]

[0162] Table 1 reveals three key observations regarding object detection results. First, the proposed MSR-FCPN model comprehensively outperforms other YOLO series models in key metrics. Taking the mAP@0.5-0.95 metric as an example, MSR-FCPN achieves 79.3%, which is 10.6%, 15.4%, and 7.4% higher than Gold-YOLO (71.7%), YOLOv10 (68.9%), and YOLOv9 (73.9%), respectively, fully demonstrating the effectiveness of its method. These results validate the effectiveness of the MSR-FCPN method. Second, MSR-FCPN demonstrates superior parameter efficiency. Although its parameter count is 4.3M, higher than YOLOv5's 1.9M and YOLOv9's 2.0M, it shows a significant advantage in parameter efficiency compared to models such as Gold-YOLO (5.6M) and YOLO-MS (5.6M). For example, compared to YOLO-MS, with the same number of parameters, mAP@0.5-0.95 is improved by 6.4%. This shows that MSR-FCPN achieves a significant improvement in detection performance with limited parameter growth, and comprehensively surpasses other models in both parameter efficiency and accuracy. Thirdly, MSR-FCPN has significant advantages in small target and occluded target detection. Its P and R reach 94.9% and 94.9% respectively, both higher than other YOLO series models. Compared to Gold-YOLO, P and R are improved by 6.0% and 1.8% respectively; compared to YOLOv10, P and R are improved by 14.9% and 11.8% respectively. These data fully demonstrate that MSR-FCPN, through multi-scale reconstruction convolution (MSRC) and feature compensation modules, can more effectively capture the detailed features of dead targets and recover the features of occluded parts of dead targets in occluded situations, thus achieving better results in dead target detection tasks. Compared to YOLOv11, MSR-FCPN outperforms YOLOv11 across all key metrics. This is because its multi-scale feature extraction and feature compensation mechanisms can more comprehensively acquire feature information from images and effectively prevent feature loss during cross-layer transmission and interaction. This stepwise fusion approach is particularly suitable for solving occlusion problems, as occlusion can lead to the loss or incompleteness of some features of dead chickens. By segmenting and recombining features, and fully considering semantic gaps when comprehensively considering the interaction information between feature layers, MSR-FCPN can more effectively improve the detection performance of dead chicken targets, thereby achieving more accurate dead chicken detection in multi-layer caged broiler environments.

[0163] ablation experiment

[0164] MSR-ELAN and FCPFN each make unique contributions to the performance of the proposed MSR-FCPN model in the dead chicken detection task. As shown in Table 2, we conducted ablation experiments to explore how the two modules affect the results of the MSR-FCPN model. In the MSR-FCPN model structure, MSR-ELAN, as the core computational unit, has a decisive impact on the overall model performance. As can be seen from Table 2, when only MSR-ELAN is used, all indicators show significant improvements to varying degrees. For example, compared to YOLOv11, P increases from 89.1% to 93.7%, R from 94.5% to 94.8%, mAP@0.5 from 93.6% to 95.5%, and mAP@0.5-0.95 from 73.5% to 77.2%. These improvements indicate that MSR-ELAN has significant advantages in feature extraction, efficiently acquiring feature information at different scales in images, laying a solid foundation for subsequent object detection tasks. When used alone, FCPFN shows improvements over YOLOv11 in various metrics. For example, P increases from 89.1% to 92.1%, R from 94.5% to 94.1%, mAP@0.5 from 93.6% to 94.4%, and mAP@0.5-0.95 from 73.5% to 74.5%. Although its individual performance is not as good as MSR-ELAN, when used in conjunction with MSR-ELAN, the two form a complete MSR-FCPN model, further improving all metrics. This strongly demonstrates that FCPFN can effectively compensate for and fuse feature information at different levels, reduce semantic gaps, and prevent the loss of feature information during transmission and interaction, thereby significantly enhancing the model's ability to detect dead chicken targets.

[0165] The MSR-FCPN model outperforms MSR-ELAN and FCPFN individually across all evaluation metrics. Particularly noteworthy are the Probability and Reduction (P / R) scores, where the complete model achieves 94.9%, significantly higher than both MSR-ELAN (P: 93.7%, R: 94.8%) and FCPFN (P: 92.1%, R: 94.1%). This result demonstrates that MSR-FCPN, through the synergistic operation of multi-scale feature extraction and feature compensation mechanisms, can more effectively capture the detailed features of dead chicken targets and recover the features of occluded parts of the dead chicken target even under occlusion conditions, thus achieving optimal performance in dead chicken detection tasks.

[0166] Table 2 Ablation experiments on different modules

[0167]

[0168] Visualization results

[0169] To further enhance the effectiveness of the MSR-FCPN model for target detection of dead chickens in multi-layered cage environments, we selected representative images for target detection. Figure 9 The image shows the target detection results of an occluded image after processing by the MSR-FCPN model and the traditional YOLO model. The second and third columns represent the detection results of the standard YOLO model and the MSR-FCPN model, respectively. The targets marked in red represent occluded dead targets that were not detected by the YOLO model. Figure 5 It is evident that in complex, severely occluded scenes, the YOLO model missed some targets, while the MSR-FCPN model successfully detected those missed by the YOLO algorithm. The main reason the MSR-FCPN model can identify the missing targets is that its Multi-Scale Reconstruction Convolution (MSRC) uses convolutional kernels of various sizes to process the input feature map in parallel. This allows MSR-FCPN to comprehensively acquire feature information at different scales in the image, accurately identifying targets even when they are occluded through captured local features and contextual information. Furthermore, the feature compensation and fusion mechanism enables MSR-FCPN to better handle the feature information of occluded targets; even if some features are occluded, it can accurately identify targets by compensating for and fusing other relevant features.

[0170] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0171] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A dead-chicken detection method based on multi-scale reconstruction layer aggregation and feature compensation progressive fusion, characterized in that, include: Acquire a target detection image, which includes dead chickens and live chickens; The target detection image is input into a preset YOLO model, which outputs a dead chicken detection result. The preset YOLO model includes a backbone feature extraction network, a feature fusion module, and a detection output head. The backbone feature extraction network is a preset multi-scale reconstruction efficient layer aggregation MSR-ELAN network, and the feature fusion module is a preset feature compensation progressive fusion pyramid FCPPN network. The preset multi-scale reconstruction efficient layer aggregation MSR-ELAN network includes four layers, and each layer includes a multi-scale reconstruction extraction block. The step of inputting the target detection image into a preset YOLO model and outputting the dead chicken detection result includes: The target detection image is input into the preset multi-scale reconstruction high-efficiency layer aggregation MSR-ELAN network, which outputs feature maps of multiple sizes. The feature maps of the multiple sizes are input into the preset feature-compensated progressive fusion pyramid FPFN network, and multiple fused feature maps are output. The multiple fused feature maps are input into the detection output head, and the dead chicken detection result is output. The process involves inputting the target detection image into the preset multi-scale reconstruction high-efficiency layer aggregation MSR-ELAN network, which outputs feature maps of multiple sizes, including: The target detection image is processed by standard convolution to generate a basic feature map; The basic feature map is input into the first-level multi-scale reconstruction extraction block, and the first feature map is output. The first feature map is processed by standard convolution to output the first convolutional feature map; After adjusting the number of channels of the first convolutional feature map through standard convolution, the number of channels of the first convolutional feature map is adjusted to C1; The first feature map is input into the second-level multi-scale reconstruction extraction block, and the second feature map is output. The second feature map is processed by standard convolution to output the second convolutional feature map; After adjusting the number of channels of the second convolutional feature map through standard convolution, the number of channels of the second convolutional feature map is adjusted to C1+C2, and the second convolutional feature map is segmented according to the number of channels C1+C2 to generate the second convolutional feature map C1 and the second convolutional feature map C2. The second feature map is input into the third-level multi-scale reconstruction extraction block, and the third feature map is output. The third feature map is processed by standard convolution to output a third convolutional feature map; After adjusting the number of channels of the third convolutional feature map through standard convolution, the number of channels of the third convolutional feature map is adjusted to C1+C2+C3, and the third convolutional feature map is segmented according to the number of channels C1+C2+C3 to generate third convolutional feature map C1, third convolutional feature map C2 and third convolutional feature map C3; The third feature map is input into the fourth-level multi-scale reconstruction extraction block, and the fourth feature map is output. The fourth feature map is processed by standard convolution to output a fourth convolutional feature map; After adjusting the number of channels of the fourth convolutional feature map through standard convolution, the number of channels of the fourth convolutional feature map is adjusted to C1+C2+C3+C4, and the fourth convolutional feature map is segmented according to the number of channels C1+C2+C3+C4 to generate fourth convolutional feature maps C1, C2, C3 and C4. In the first level, the second convolutional feature map C1, the third convolutional feature map C1, and the fourth convolutional feature map C1 are upsampled and then combined with the first convolutional feature map to generate a feature map of the first size. In the second level, the third convolutional feature map C2 and the fourth convolutional feature map C2 are upsampled and then combined with the second convolutional feature map C2 to generate a feature map of the second size. In the third level, the fourth convolutional feature map C3 is upsampled and then combined with the third convolutional feature map C3 to generate a feature map of the third size. In the fourth level, the fourth convolutional feature map is determined as a feature map of the fourth size; The process involves inputting the feature maps of the multiple sizes into the preset feature-compensated progressive fusion pyramid (FCPFN) network, and outputting multiple fused feature maps, including: In order of size, feature maps of every two adjacent pairs are subjected to feature compensation processing and then standard convolution processing to generate compensated feature maps, including: Feature compensation processing is performed based on the feature map of the first size and the feature map of the second size, followed by standard convolution processing to generate a first compensated feature map; Feature compensation processing is performed based on the feature map of the second size and the feature map of the third size, followed by standard convolution processing to generate a second compensated feature map; Feature compensation processing is performed based on the feature map of the third size and the feature map of the fourth size, followed by standard convolution processing to generate a third compensated feature map; Perform standard convolution on the feature map of the fourth size to generate a fourth compensated feature map; After upsampling the second compensation feature map, it is adaptively fused with the first compensation feature map and then subjected to standard convolution to generate a fused feature map A1. After downsampling the first compensation feature map, it is adaptively fused with the second compensation feature map and then subjected to standard convolution to generate a fused feature map B1. After downsampling the fused feature map B1 and the fused feature map A1, they are adaptively fused with the third compensation feature map and then convolved to generate the fused feature map C1. After upsampling the fused feature map B1 and the third compensation feature map, they are adaptively fused with the fused feature map A1 and then convolved to generate the fused feature map A2. After downsampling and upsampling of the fused feature map A1 and the third compensation feature map respectively, they are adaptively fused with the fused feature map B1 and then convolved to generate the fused feature map B2. After upsampling the fused feature map B2, the fused feature map C1, and the fourth compensation feature map, adaptive fusion is performed with the fused feature map A2 to generate a first-size fused feature map; After downsampling the fused feature map A2, upsampling the fused feature map C1 and the fourth compensation feature map, and adaptively fusing them with the fused feature map B2, a second-size fused feature map is generated. After downsampling the fused feature map A2 and the fused feature map B2, and upsampling the fourth compensation feature map, adaptive fusion is performed with the fused feature map C1 to generate a third-size fused feature map; After downsampling the fused feature map A2, the fused feature map B2, and the fused feature map C1, adaptive fusion is performed with the fourth compensation feature map to generate a fourth-size fused feature map.

2. The method as described in claim 1, characterized in that, The step of inputting the basic feature map into the first-level multi-scale reconstruction extraction block and outputting the first feature map includes: After performing convolution processing on the basic feature map, it is divided into a first basic feature sub-map and a second basic feature sub-map. The first basic feature sub-image is convolved to generate the first basic convolutional feature sub-image. The second basic feature sub-map is subjected to multi-scale reconstruction convolution processing and repeated multiple times to generate the second basic convolution feature sub-map. The first basic convolutional feature sub-image and the second basic convolutional feature sub-image are sequentially concatenated and convolved to generate the first feature map.

3. The method as described in claim 2, characterized in that, The step of performing multi-scale reconstruction convolution processing on the second basic feature sub-image includes: The second basic feature sub-map is sequentially subjected to convolution, batch normalization, and non-linear activation by the ReLU6 activation function to generate an extended activation feature map. The extended activation feature map is simultaneously subjected to multiple structural reconstruction convolutional SCConv processes to generate multiple reconstructed feature maps, wherein the scale of the convolutional processing in each SCConv process is different; The multiple reconstructed feature maps are sequentially batch normalized, nonlinearly activated by the ReLU6 activation function, fused, and then subjected to channel shuffling and convolution.

4. The method as described in claim 3, characterized in that, The process of simultaneously performing multiple structural reconstruction convolutional SCConv processes on the expanded activation feature map generates multiple reconstructed feature maps, including: Each time, the extended activation feature map is processed sequentially by spatial reconstruction SRU, channel reconstruction CRU, and convolution to generate the reconstructed feature map.

5. The method as described in claim 4, characterized in that, The process of performing feature compensation processing on every two adjacent feature maps in order of size, followed by standard convolution processing to generate compensated feature maps, includes: After performing convolution and upsampling on the first feature map, the difference between it and the second feature map after convolution is calculated to obtain the lost key information. The lost key information is added to the second feature map after convolution to generate the compensation feature map.

6. The method as described in claim 5, characterized in that, The adaptive fusion includes: Assign corresponding weights to the multiple feature maps for adaptive fusion; Based on the weights corresponding to each feature map, the multiple feature maps that are adaptively fused are weighted and fused.

Images