Physical guiding channel intelligent dynamic self-pruning neural network and method

Abstract

The application discloses a kind of physical guiding channel intelligent dynamic self-pruning neural network and method, including main network, neck network and detection head. Wherein main network contains multiple front-end module, configurable multi-scale feature extraction module CMOC based on open circuit mechanism, physical information guide's self-pruning channel attention module PSCA and multiple rear-end module. The CMOC module is introduced in the main network in the application, and the start-stop state of different receptive field convolution branch is dynamically regulated using circuit breaker type switch, to realize the flexible on-demand configuration of multi-scale feature extraction topology. At the same time, the PSCA module is embedded in the main network and the neck network, the spatial weight is generated using soft pooling operation, and the dynamic pruning of channel level is driven through the learning multi-layer perception machine, to efficiently suppress redundancy while retaining key features. Through collaborative work, the simultaneous optimization of detection accuracy and computational efficiency is realized.

Description

Technical Field

[0001] This application belongs to the field of deep learning image processing technology, specifically relating to a physical guidance channel intelligent dynamic self-pruning neural network and method. Background Technology

[0002] In recent years, deep learning technology has achieved remarkable success in the field of computer vision, especially in object detection tasks. To improve model performance, industry research has mainly focused on two core directions: first, enhancing feature representation capabilities to improve detection accuracy; and second, optimizing model structure to reduce computational complexity.

[0003] In feature extraction, multi-scale feature fusion is a widely adopted and effective strategy. For example, by integrating convolutional kernels of different sizes to construct parallel or serial structures, features with diverse receptive fields can be captured, thereby improving the model's ability to detect targets at different scales. Furthermore, attention mechanisms, especially channel attention modules, have been shown to effectively improve model performance by recalibrating channel feature responses to allow the network to adaptively focus on feature channels with richer information. Among these, squeeze-and-excitation (SE) networks and their variants are representative works in this field, typically relying on global average pooling (GAP) and fully connected layers to generate channel weights.

[0004] However, existing technologies still have several inherent limitations that restrict further optimization of model performance:

[0005] First, in the feature extraction stage, existing multi-scale modules (such as Inception-type structures) typically have static and fixed topologies. This means that once the model is trained, the scale and combination of extracted features cannot be adaptively adjusted according to the specific content of the input samples. This structural rigidity results in the model lacking sufficient flexibility for complex and varied detection scenarios, limiting its generalization ability in feature representation. Furthermore, integrating multiple parallel branches usually introduces a significant increase in parameters and computational cost.

[0006] Secondly, the pooling operations relied upon by mainstream channel attention mechanisms in the feature selection stage have inherent limitations. Both Global Max Pooling (GMP) and Global Average Pooling (GAP) are preprocessing methods that degenerate and aggregate spatial information. GMP retains only the single strongest response point, while GAP uniformly weights all spatial locations. Both methods inevitably lead to the loss or smoothing of fine-grained spatial information, failing to accurately characterize the differences in the contribution of different spatial locations in the feature map to the final task. The channel weights generated by this coarse aggregation method have a ceiling in discriminability and accuracy, thus affecting the effectiveness of subsequent feature recalibration.

[0007] Therefore, existing solutions struggle to achieve intelligent, precise, and dynamic focusing on key features while maintaining high operational efficiency, becoming a technical bottleneck hindering further improvements in target detection performance. Summary of the Invention

[0008] The purpose of this application is to overcome the aforementioned deficiencies of the prior art and provide a physically guided intelligent dynamic self-pruning neural network and method. This network, by introducing a dynamically configurable multi-scale feature extraction module and a refined channel attention mechanism based on soft pooling, aims to achieve adaptive optimization of the model structure and feature weights, thereby simultaneously improving the accuracy and efficiency of target detection.

[0009] To achieve the above technical objectives, this application specifically adopts the following technical solution:

[0010] In one aspect of this application, a physically guided channel intelligent dynamic self-pruning neural network is provided for image detection, comprising:

[0011] The backbone network is used for hierarchical feature extraction from the input image;

[0012] The neck network, with its input end connected to the output end of the backbone network, is used to receive and fuse shallow and deep features from different depths of the backbone network.

[0013] The detection head, with its input end connected to the output end of the neck network, is used to predict the target category and location based on the fused multi-scale feature map;

[0014] The backbone network includes:

[0015] Multiple front-end feature extraction modules are used to perform preliminary downsampling and feature mapping on the input image;

[0016] The Configurable Multi-Scale Feature Extraction Module (CMOC) based on the open-path mechanism has its input connected to the output of the front-end feature extraction module. It is used to receive the preliminary feature map and extract and fuse features from different receptive fields through multiple parallel convolutional paths that are independently controlled by adjustable parameters.

[0017] The Physical Information Guided Self-Pruning Channel Attention Module (PSCA) has its input connected to the output of the CMOC module. It is used to perform soft pooling on the fused multi-scale feature map to calculate spatial weights and generate channel weights through a multilayer perceptron. Based on this, the feature channels are dynamically pruned to retain key features.

[0018] Multiple back-end feature extraction and fusion modules, whose inputs are connected to the output of the PSCA module, are used to further process the pruned and optimized features and output multi-scale deep features to the neck network.

[0019] In one implementation, the CMOC module includes:

[0020] Multiple parallel convolutional paths, each corresponding to a convolutional kernel of a specific size, are interconnected at their input ends and jointly receive the input feature map;

[0021] The channel stitching layer, whose input is connected to the output of all activated parallel convolutional paths, is used to stitch the feature maps output by each path along the channel dimension.

[0022] Each convolutional path is equipped with a switch unit that can be controlled by external parameters. When the switch unit is in the first state, the path is activated and its output is transmitted to the channel splicing layer. When the switch unit is in the second state, the path is disabled.

[0023] In one implementation, the multiple parallel convolutional pathways include convolutional layers with kernel sizes of 3×3, 5×5, 7×7, and 9×9.

[0024] In one implementation, the PSCA module includes:

[0025] The soft pooling unit, whose input is connected to the output of the CMOC module, is used to calculate the probability weight of each spatial location in the feature map through the Softmax function, and generate the initial channel weight vector by weighted summation.

[0026] A multilayer perceptron (MLP) has its input connected to the output of the soft pooling unit, and is used to perform a nonlinear transformation on the initial channel weight vector to generate the final channel importance weight vector.

[0027] The channel weighting and pruning unit has two input terminals. Its first input terminal receives the output feature map directly from the CMOC module, and its second input terminal is connected to the output terminal of the multilayer perceptron (MLP) to receive the channel importance weight vector, which is used to multiply the feature map of each channel with the corresponding weight and suppress redundant channels with weights below a threshold.

[0028] In one implementation, the soft pooling unit calculates the first channel weight in the initial channel weight vector. m The weight of each channel x m for:

[0029]

[0030] in, B m Indicates the first m Spatial weight matrix for each channel, each elementb i,j The spatial weights are calculated using the Softmax function and represent the positions in the feature map. i,j The importance of ); A m Indicates the first m The original eigenvalue matrix of each channel, each element a i,j It is the first m Each channel is located at ( i, j The eigenvalue (activation value) at position ) R This represents the spatial dimension of the entire feature map.

[0031] In one implementation, the multilayer perceptron comprises three fully connected layers, wherein the activation function is the SiLU function.

[0032] In one implementation, the backend feature extraction module in the backbone network includes cascaded C3k2 modules;

[0033] The backbone network is also connected to a Spatial Pyramid Fast Pooling (SPPF) module and a Cross-Level Pyramid Slice Attention (C2PSA) module at its end. The input of the SPPF module is connected to the output of the back-end feature extraction module, and the input of the C2PSA module is connected to the output of the SPPF module. The output of the C2PSA module is connected to the neck network.

[0034] In one implementation, the neck network increases the size of the deep feature map through an upsampling operation and fuses it with the feature map from the shallow layer of the backbone network through a channel splicing operation; the neck network further integrates at least one of the CMOC modules and / or at least one of the PSCA modules for multi-scale extraction and dynamic optimization of features during the fusion process.

[0035] In one implementation, the detection head includes three branches, which are respectively connected to feature maps of three different scales output by the neck network, for target detection at three resolutions: 80×80, 40×40, and 20×20.

[0036] In another aspect of this application, a target detection method based on a neural network is provided, using the physically guided channel intelligent dynamic self-pruning neural network as described above, the method comprising:

[0037] The image to be detected is input into the backbone network of the neural network;

[0038] Configurable multi-scale feature fusion is performed through the CMOC module in the backbone network;

[0039] The feature channels are dynamically pruned using the PSCA module in the backbone network.

[0040] The neck network is fused with multi-scale features extracted from the backbone network;

[0041] The detection head outputs the final target category and location prediction results.

[0042] The beneficial effects of this application are as follows:

[0043] 1) Significantly Improved Detection Accuracy: This application employs a configurable multi-scale feature extraction module (CMOC) based on an open-path mechanism. The model can dynamically activate convolutional pathways with different receptive fields according to the specific task and input content, achieving adaptive, multi-scale refined feature extraction and greatly enhancing the ability to identify targets with significant scale differences. Simultaneously, a physically-guided self-pruning channel attention module (PSCA) replaces traditional pooling operations, utilizing soft pooling to more precisely preserve spatial information and generate more discriminative channel weights, thereby accurately focusing on key feature channels. Experiments demonstrate that the model in this application achieves optimal performance in core metrics such as mAP50, mAP50-95, and average recall.

[0044] 2) Model Efficiency and Generalization Optimization: The configurable nature of the CMOC module allows for a more flexible model topology, enabling the most necessary computational paths to handle different scenarios and avoiding redundant computations. The dynamic pruning mechanism of the PSCA module automatically suppresses redundant channels with low contribution, further reducing the model's computational complexity and overfitting risk. Experiments have demonstrated that, in comparisons with models of similar or even greater parameter counts, this application achieves higher accuracy with lower computational cost, effectively balancing accuracy and efficiency.

[0045] 3) Strong structural flexibility and scalability: The CMOC and PSCA modules proposed in this application are designed as plug-and-play structures, which can be embedded into different layers of existing neural networks (such as the YOLO series) without changing the overall framework of the backbone network. The modular design gives the model good flexibility and scalability, making it easy to customize deployment for specific application scenarios.

[0046] In summary, this application effectively solves the problems of rigid feature extraction structure and loss of feature selection information in traditional methods by leveraging the synergistic effect of the two core innovative modules, CMOC and PSCA, and achieves simultaneous optimization of detection accuracy and operating efficiency. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the structure of the intelligent dynamic self-pruning neural network for the physical guidance channel of this application;

[0048] Figure 2 This is a schematic diagram of the internal structure of each module adopted from YOLOv11 in this application;

[0049] Figure 3 This is a schematic diagram of the internal structure of the CMOC module in this application;

[0050] Figure 4 This is a schematic diagram of the internal structure of the PSCA module in this application;

[0051] Figure 5 This is the target detection result of the embodiment of this application. Detailed Implementation

[0052] The technical solution of this application will be clearly and completely described below with reference to specific embodiments. However, those skilled in the art will understand that the embodiments described below are only some embodiments of this application, not all embodiments, and are only used to illustrate this application, and should not be regarded as limiting the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0053] In one embodiment of this application, reference is made to Figure 1 As shown, a physically guided intelligent dynamic self-pruning neural network is provided for image detection, including a backbone network for hierarchical feature extraction of the input image; a neck network connected to the output of the backbone network for receiving and fusing shallow and deep features from different depths of the backbone network; and a detection head connected to the output of the neck network for predicting the target category and location based on the fused multi-scale feature map.

[0054] In this embodiment, the model inherits the three-layer basic architecture of YOLOv11—backbone, neck, and detection head—and incorporates the Configurable Multi-Scale Feature Extraction Module (CMOC) based on an open-path mechanism and the Physical Information Guided Self-Pruning Channel Attention Mechanism (CMOC), significantly improving model performance. The Configurable Multi-Scale Feature Extraction Module (CMOC) based on an open-path mechanism is introduced into the backbone network. It dynamically controls the start and stop of convolutional branches in different receptive fields using a circuit breaker-like switch, enabling flexible on-demand configuration of the network topology. Secondly, the Physical Information Guided Self-Pruning Channel Attention Mechanism (CMOC) is embedded in the backbone and neck. This module utilizes soft pooling to generate spatial weights and drives dynamic channel pruning through a learnable multilayer perceptron, efficiently eliminating redundancy while preserving key features. Together, these features enable the model to simultaneously optimize accuracy and computational efficiency while maintaining the efficient detection framework of the YOLO series.

[0055] The backbone network is responsible for extracting hierarchical features from the input image, gradually compressing the spatial dimension and expanding the number of channels to capture multi-scale features.

[0056] Specifically, the backbone network includes a front-end feature extraction module, a configurable multi-scale feature extraction module CMOC based on an open-path mechanism, a physical information-guided self-pruning channel attention module PSCA, and multiple back-end feature extraction and fusion modules.

[0057] The front-end feature extraction module is used to perform preliminary downsampling and feature mapping on the input image. Its purpose is to gradually reduce the spatial size (width and height) of the feature map while increasing its number of channels to capture multi-level feature representations from low to medium levels, and to provide high-quality, multi-scale input features for subsequent core processing modules (such as CMOC and PSCA).

[0058] In some embodiments, reference is made to Figure 1 As shown, the front-end feature extraction module consists of four CBS modules and three C3k2 modules connected in a specific order. The module progressively reduces the image size from 640×640 to 40×40 using the four downsampling CBS modules, and performs in-depth feature learning and fusion at three key scales: 160×160, 80×80, and 40×40 using the three C3k2 modules. Finally, the module outputs one or more feature maps that have undergone preliminary downsampling and deep feature mapping. These feature maps are then used as input to a subsequent Configurable Multi-Scale Feature Extraction (CMOC) module based on an open-path mechanism for further processing. The specific hierarchical structure and data processing procedure are as follows:

[0059] For example, an input RGB image with dimensions of 640×640×3 is processed by the first CBS module. (See reference...) Figure 2 As shown, the CBS module consists of three layers: a 2D convolutional layer, a batch normalization layer, and a SiLU activation function layer. The 2D convolutional layer within the CBS module is configured with kernel size (k) = 3, stride (s) = 2, and padding (p) = 1. This operation downsamples the feature map space to half its original dimensions, i.e., 320×320. The number of output channels (c) is set to a base value (e.g., 32 or 64). After processing by the batch normalization layer and the SiLU activation function layer, the output feature map size is 320×320×C1 (C1 is the number of output channels).

[0060] The obtained 320×320×C1 feature map is further downsampled to 160×160 by the second CBS module. The number of output channels is usually increased, for example, C2=2×C1, and the output feature map size is 160×160×C2.

[0061] A 160×160×C2 feature map is fed into the first C3k2 module. The C3k2 module first adjusts the channel count using a CBS (typically k=1, s=1), then splits the feature map into two parts. One part undergoes deep feature processing through multiple sequences of Bottleneck structures (each containing two CBSs with k=3, s=1), while the other part is directly connected via short-circuiting. Finally, the two feature parts are merged through a concatenation operation and then integrated using another CBS (k=1, s=1), resulting in an output feature map with the same size of 160×160×C3. Understandably, this module does not change the spatial size of the feature map but significantly enhances its feature representation capability and receptive field.

[0062] The feature map 160×160×C3 is further processed by the third CBS module with downsampling convolution of k=3, s=2, p=1, reducing the feature map size to 80×80. The number of output channels continues to increase, for example, C4=2×C3, and the output feature map is 80×80×C4.

[0063] The feature map 80×80×C4 is further refined and extracted through the second C3k2 module, resulting in the output feature map 80×80×C5.

[0064] The feature map 80×80×C5 is subjected to a final standard downsampling with k=3, s=2, p=1 by the fourth CBS module, reducing the feature map size to 40×40 and increasing the number of output channels to C6, resulting in an output feature map of 40×40×C6.

[0065] The feature map 40×40×C6 undergoes final feature deepening and refinement through the third C3k2 module, ensuring that the features passed to subsequent modules are highly refined, and outputs feature map 40×40×C7.

[0066] In some embodiments, reference is made to Figure 2 As shown, the C3k2 module consists of three CBS modules and multiple cascaded C3k bottleneck structures, achieving efficient multi-scale feature fusion and hierarchical representation learning through cross-stage partial connections. The specific structure is as follows:

[0067] The input feature map first undergoes preliminary feature extraction and standardization in a CBS module, which includes a 2D convolutional layer, a batch normalization layer, and a SiLU activation function. After a Split operation, the feature map is split into two parts. One part is directly passed to subsequent layers, retaining the original feature information; the other part enters a sub-network consisting of two cascaded C3k bottleneck structures for deep feature extraction. Finally, the features from the two branches are concatenated along the channel dimension. Each C3k structure contains two CBS modules and a residual connection. The bottleneck design reduces the number of parameters and enhances gradient flow. The cascading of multiple C3k structures further enhances the model's non-linear expressive power and feature reuse efficiency. The features processed by the C3k sub-network are concatenated with the previously directly passed features along the channel dimension, achieving cross-stage partial feature fusion. This design retains shallow details while incorporating deep semantic features, effectively improving the model's multi-scale perception capability. The concatenated features are then integrated and dimensionality-reduced in a final CBS module, ultimately outputting a feature map with rich multi-scale contextual information.

[0068] The input end of the CMOC module is connected to the output end of the front-end feature extraction module. It is used to receive the preliminary feature map and extract and fuse features from different receptive fields through multiple parallel convolutional paths controlled independently by adjustable parameters.

[0069] The CMOC module includes multiple parallel convolutional paths and a channel concatenation layer. Each parallel convolutional path corresponds to a convolutional kernel of a specific size. The inputs of the multiple parallel convolutional paths are interconnected and jointly receive the input feature map. The input of the channel concatenation layer is connected to the output of all activated parallel convolutional paths, and is used to concatenate the feature maps output by each path along the channel dimension. Each convolutional path has a switch unit that can be controlled by external parameters. When the switch unit is in a first state, the path is activated, and its output is transmitted to the channel concatenation layer; when the switch unit is in a second state, the path is disabled.

[0070] The CMOC module, through its unique parallel structure and independently controllable switching mechanism, enables flexible configuration and on-demand fusion of multi-scale feature extraction, overcoming the shortcomings of fixed structure and computational redundancy in traditional multi-scale modules, and effectively improving model detection performance and efficiency.

[0071] In some embodiments, reference is made to Figure 3As shown, the multiple parallel convolutional paths include four independent convolutional layers with kernel sizes (k) of 3×3, 5×5, 7×7, and 9×9, respectively. To ensure that the output feature maps after processing all branches have the same spatial dimensions (H, W) for subsequent splicing and fusion, the stride (s) of each convolutional layer is set to 2, and a specially calculated padding (p) value is assigned: p=1 for the convolutional layer with k=3; p=2 for the convolutional layer with k=5; p=3 for the convolutional layer with k=7; and p=4 for the convolutional layer with k=9.

[0072] The number of output channels in each convolutional layer is consistent with the number of channels in the input feature map (i.e., C). The spatial size of the output feature map is calculated using the following formula, becoming half the input size (e.g., if the input is 40×40, then the output of each channel is 20×20):

[0073]

[0074] in, Indicates the size of the input feature map. This represents the size of the output feature map.

[0075] In some embodiments, a switching unit is provided on each convolutional path. The state of this switching unit is independently controlled by an external, learnable, or pre-configured set of binary parameters {K1, K2, K3, K4}, where each parameter K... i ∈{0,1} corresponds to one convolutional path.

[0076] When the switching parameter K of a certain path i When =1 (first state), the switch unit is in the closed state, the convolution path is activated, its convolution operation is performed, and the resulting output feature map will be passed to the subsequent channel splicing layer.

[0077] When the switching parameter K of a certain path i When =0 (first state), the switch unit is in the off state, this convolution path is disabled, its convolution operation is not performed in this forward propagation, and no output is passed to the splicing layer.

[0078] The "open-path mechanism" control method endows the CMOC module with extremely high flexibility and configurability. Users or the model itself can dynamically select which feature extraction pathways to enable based on specific task requirements, computational resource constraints, or prior knowledge of the importance of targets at different scales, thereby achieving dynamic reconstruction of the module's topology.

[0079] In some embodiments, the input of the channel stitching layer is connected to the output of all activated (i.e., switched in the first state) parallel convolutional paths. The channel stitching layer stitches together the feature maps output by all activated paths, for example, all of which are 20×20×C in size, along the channel dimension.

[0080] Assuming N pathways are activated (N is an integer between 1 and 4), the size of the fused output feature map is 20×20×(N×C). This output feature map contains fused features extracted from multiple receptive fields corresponding to the activated pathways, and is then sent to the next module (such as the PSCA module) for further processing.

[0081] The PSCA module's input is connected to the CMOC module's output. It performs soft pooling on the fused multi-scale feature map to calculate spatial weights and generates channel weights using a multilayer perceptron (MLP). Based on this, it dynamically prunes the feature channels to preserve key features. The PSCA module includes a soft pooling unit, a MLP, and a channel weighting and pruning unit.

[0082] In some embodiments, the soft pooling unit calculates the probability weight of each spatial location in the feature map using the Softmax function, and generates an initial channel weight vector by weighted summation.

[0083] In some embodiments, reference is made to Figure 4 As shown, the initial channel weight matrix is ​​first calculated using the Softmax function, assigning a weight value to each pixel in the feature map. Assume the original feature values ​​are... Then the initial spatial weights It can be represented as:

[0084]

[0085] in, The spatial dimensions of the entire feature map, It is the Euler number. a i,j It is a certain channel in position ( i,j The eigenvalue (activation value) at position ). Indicates all spatial locations within the current channel ( k,l The indexed eigenvalues ​​of ) The summation is performed, and the result is a normalization factor that ensures that the sum of the weights of all points in the channel is 1.

[0086] Therefore, in the initial channel weight matrix, the first... m The weight of each channel x m for:

[0087]

[0088] in, B m Indicates the first m Spatial weight matrix for each channel, each element b i,j The spatial weights are calculated using the Softmax function and represent the positions in the feature map. i,j The importance of ); A m Indicates the first m The original eigenvalue matrix of each channel, each element a i,j It is the first m Each channel is located at ( i,j The eigenvalue (activation value) at ().

[0089] This method avoids the information loss that may be caused by traditional max pooling or average pooling, and can more meticulously preserve discriminative spatial information in the feature map.

[0090] The input of the multilayer perceptron (MLP) is connected to the output of the soft pooling unit, and is used to perform a nonlinear transformation on the initial channel weight vector to generate the final channel importance weight vector.

[0091] In some embodiments, the multilayer perceptron (MLP) consists of three fully connected layers, where the activation function is the SiLU function, which is used to perform a nonlinear transformation on the initial weights to generate the final channel importance weight vector. y Its mathematical expression is:

[0092]

[0093] in, It is the weight matrix of the Lth layer. It is a bias vector. The SiLU activation function is expressed mathematically as follows:

[0094]

[0095] The role of MLP is to further refine channel weights and enhance the model's ability to model the dependencies between channels.

[0096] The channel weighting and pruning unit has two input terminals. Its first input terminal receives the output feature map from the CMOC module, and its second input terminal is connected to the output terminal of the multilayer perceptron to receive the channel importance weight vector. This unit multiplies the feature map of each channel with the corresponding channel weight to achieve channel weighting. At the same time, a learnable threshold (or preset threshold) is set to suppress or prune channels with weights below the threshold, retaining only the key channels that contribute significantly to the task, thereby achieving dynamic channel self-pruning, reducing computational redundancy, and improving model efficiency.

[0097] In some embodiments, the backend feature extraction module in the backbone network includes a C3k2 module. The C3k2 module consists of three CBS modules and multiple cascaded C3k bottleneck structures. Through cross-stage partial connections, it achieves efficient multi-scale feature fusion and hierarchical representation learning, significantly improving the model's discriminative ability and computational efficiency during feature extraction.

[0098] In some embodiments, the backbone network is further connected to the ends of a Spatial Pyramid Fast Pooling (SPPF) module for multi-scale contextual information aggregation and a Cross-Level Pyramid Slice Attention (C2PSA) module for enhancing multi-scale feature representation capabilities.

[0099] The input of the SPPF module is connected to the output of the last C3k2 module in the back-end feature extraction module, and is used to receive its output feature map. The SPPF module contains a CBS layer at both ends, and integrates multiple levels of parallel or serial max pooling operations in between. The feature maps of different receptive fields generated by these pooling operations are fused in the channel dimension through a concatenation layer to form an efficient multi-scale context feature extractor, which significantly enhances the model's robustness to changes in target scale.

[0100] In some embodiments, the structure and data flow of the SPPF module are as follows:

[0101] The input feature map first undergoes preliminary feature extraction and normalization processing through a CBS module (Conv-BatchNorm-SiLU). This CBS module contains a 2D convolutional layer, a batch normalization layer, and a SiLU activation function. Subsequently, the feature map is input in parallel to an identity connection and three sequential max pooling layer branches. The feature map from the identity connection and the feature maps output from the three sequential pooling layers are concatenated along the channel dimension. The concatenated fused feature map then passes through another CBS module for feature integration and dimensionality reduction. This module fuses and compresses multi-scale features, outputting a final feature map rich in multi-scale contextual information.

[0102] In this design, the identity connection directly preserves the original feature information, and the three sequential max-pooling layers all use the same kernel size, ensuring that the feature map spatial size after each pooling operation remains consistent with the input. The pooling layers process the feature map sequentially in a cascaded manner, with each pooling layer performing one pooling operation on the input. Due to the identical parameter settings, their output sizes are consistent. This sequential, repetitive pooling operation allows each subsequent pooling layer to have a larger effective receptive field, thereby capturing contextual information of different ranges.

[0103] The input of the C2PSA module is connected to the output of the SPPF module, and it is used to further optimize the output features rich in multi-scale contextual information. The C2PSA module is a multi-scale attention mechanism that divides the input feature map into two equal parts along the channel dimension. One part is directly propagated forward, while the other part is processed through two cascaded pyramid slice attention PSA sub-modules. Finally, the two parts of the output feature map are concatenated along the channel dimension, then integrated and dimensionality-reduced through a CBS layer, and finally sent out from its output to the subsequent neck network. This design, through cross-layer feature fusion and channel attention mechanisms, enables the model to better focus on salient features at different scales.

[0104] In some embodiments, reference is made to Figure 2 As shown, the structure and data flow of the C2PSA module are as follows:

[0105] The input feature map is first preprocessed by a CBS module, which performs initial channel dimension adjustment and feature compression. The preprocessed feature map is then split into two equal parts. One part is directly passed through identity connections, while the other part is enhanced by a deep attention branch consisting of two Pyramid Slice Attention (PSA) modules connected in series. The original features from the identity connections and the enhanced features from the deep attention branch (after two serial PSA processes) are concatenated along the channel dimension. The resulting fused feature is then passed through the CBS module for final channel adjustment and feature fusion, generating the final output of the C2PSA module. This output feature map contains both the original features with complete information and the salient features filtered and enhanced by the deep multi-scale attention mechanism.

[0106] The internal structure of each PSA module is as follows: the input feature map is uniformly divided into multiple sub-feature maps along the channel dimension, and each is fed into an independent Spatial Attention Unit (SE). Each SE unit captures spatial context information at different scales. All sub-feature maps calibrated by SEs at different scales are concatenated along the channel dimension and then integrated by a CBS module to output the final enhanced features. Specifically, each SE unit first compresses the spatial information into a channel descriptor vector using Global Average Pooling (GAP). This vector is then processed by a lightweight Multilayer Perceptron (MLP), which typically contains a dimensionality reduction layer (using ReLU activation) and a dimensionality increase layer. Finally, a weight vector representing the importance of each channel is generated using a Sigmoid activation function. This weight vector is multiplied by the original sub-feature map to complete the channel-level recalibration. Preferably, a CBS layer is set between the global average pooling (GAP) and the multilayer perceptron (MLP) of each SE unit for downsampling or spatial feature extraction. By combining convolutional kernels (or pooling operations) of different sizes with the SE mechanism, a parallel channel slice pyramid structure is constructed, enabling the PSA module to capture spatial context information at different scales simultaneously.

[0107] The backbone network of this application, through the specific combination and connection relationship of the CBS module, C3k2 module, SPPF module and C2PSA module, constitutes an efficient and robust feature extraction pipeline, providing rich and high-quality multi-scale feature representation for the subsequent neck network and detection head.

[0108] In this application, the neck network is designed to achieve multi-scale feature fusion by receiving and processing feature maps from different depths of the backbone network to aggregate rich semantic and detail information.

[0109] In some embodiments, the neck network increases the spatial size of the deep feature map through an upsampling operation. The deep feature map originates from the deep outputs at the ends of the backbone network and has a high degree of semantic abstraction but low spatial resolution. The upsampling operation is preferably implemented using interpolation algorithms (such as nearest neighbor interpolation or bilinear interpolation) or transposed convolution, with the goal of enlarging the size of the deep feature map to match the size of the target shallow feature map.

[0110] Subsequently, the neck network fuses the upsampled deep feature map with the shallow feature map from the backbone network through a channel concatenation operation. The shallow feature map originates from earlier layers in the backbone network, possessing high spatial resolution and rich detail information, but with lower semantic abstraction. The channel concatenation operation concatenates these two types of feature maps along the channel dimension, thereby generating a fused feature map that simultaneously contains deep semantic information and shallow detail information, providing a more comprehensive feature representation for subsequent object detection.

[0111] To further enhance the representation capabilities and efficiency of the neck network, the neck network further integrates at least one of the open-path-based adjustable multi-scale feature extraction modules CMOC and / or at least one of the physically-guided self-pruning soft channel attention mechanism modules PSCA.

[0112] The CMOC module is deployed in the feature fusion pathway of the neck network. It receives feature maps from the aforementioned concatenation operation or other layers as input, and performs multi-scale feature extraction and downsampling on the input features through parallel convolutional pathways with different kernel sizes (e.g., 3×3, 5×5, 7×7, 9×9). The activation and deactivation of each pathway are dynamically controlled by an adjustable parameter set {K1, K2, K3, K4}, enabling the neck network to flexibly adjust the feature extraction scale according to specific task requirements, enhancing the model's ability to perceive and fuse targets at different scales.

[0113] The PSCA module is also embedded in the neck network. It receives feature maps (e.g., output from the CMOC module or fused features) as input, calculates spatial weights through soft pooling, and generates channel importance weight vectors via a multilayer perceptron (MLP). Finally, it dynamically weights and prunes the feature channels accordingly. This process suppresses redundant channels generated during feature fusion, highlights and retains key feature information for the detection task, thereby optimizing computational efficiency while maintaining or even improving model accuracy.

[0114] In some implementations, refer to Figure 1 As shown, the neck network includes: two upsampling layers for enlarging the feature map size to facilitate fusion with shallow features; four C3k2 modules for feature extraction and cross-stage partial connections to enhance multi-scale feature fusion capabilities; four channel stitching modules for stitching feature maps from different sources along the channel dimension; one CBS module containing convolution, batch normalization, and SiLU activation functions for feature transformation; one configurable multi-scale feature extraction module CMOC based on an open-path mechanism, supporting dynamic path selection; and one PSCA module, a physically-guided self-pruning channel attention mechanism for dynamic channel weight adjustment and pruning.

[0115] Specifically, the processing flow is as follows:

[0116] Multi-scale feature maps from the backbone network, such as those with sizes of 80×80, 40×40, and 20×20, enter the neck network. The 20×20 feature map is upsampled and enlarged to 40×40, then concatenated with the 40×40 feature map from the backbone network. The concatenated 40×40 feature map is then processed by the C3k2 module for further feature extraction. This processed 40×40 feature map is then upsampled again to 80×80 and concatenated with the 80×80 feature map from the backbone network. The 80×80 feature map is then processed by the C3k2 module, downsampled to 40×40 by the CBS module, and concatenated with the previously processed 40×40 feature map. This concatenated 40×40 feature map is then processed by the C3k2 module, downsampled to 20×20 by the CBS module, and concatenated with the 20×20 feature map from the backbone. The concatenated 20×20 feature map is then processed by the CMOC module and then by the CMOC module to further enhance multi-scale features and perform channel pruning. Finally, the processed multi-scale feature maps (80×80, 40×40, and 20×20) are fed into the detection head for target detection.

[0117] In some embodiments, the detection head employs a multi-scale detection architecture comprising three independent branches. The input of each branch is connected to feature maps of three different scales ultimately output by the neck network.

[0118] The three feature maps have different spatial resolutions: high-resolution (80×80 pixels), medium-resolution (40×40 pixels), and low-resolution (20×20 pixels). Each feature map is directed to a corresponding detection branch for processing.

[0119] Each detection branch has the same internal structure, containing two convolutional block structure (CBS) modules and a final two-dimensional convolutional layer. The CBS modules sequentially perform convolution, batch normalization, and SiLU activation function operations to further refine and integrate the input features, while the final two-dimensional convolutional layer is responsible for outputting the specific detection result at that scale.

[0120] The three branches work in parallel, each performing target detection at the scale of the feature map it is connected to: the branch connected to the 80×80 high-resolution feature map focuses on detecting small-scale targets in the image; the branch connected to the 40×40 medium-resolution feature map focuses on detecting medium-scale targets in the image; and the branch connected to the 20×20 low-resolution feature map focuses on detecting large-scale targets in the image.

[0121] The architecture of parallel detection using multi-scale feature maps enables the neural network to simultaneously and efficiently perceive and accurately locate targets of different sizes in an image, significantly improving the model's overall detection performance and scale adaptability in complex scenes.

[0122] In some embodiments, reference is made to Figure 2 As shown, the structure and connection relationship of the detection head are as follows:

[0123] The multi-scale feature maps output by the neck network are input to three independent, structurally identical detection heads, which process feature maps of different spatial sizes, such as 80×80 (corresponding to shallow, high-resolution features, which are beneficial for detecting small targets), 40×40 (medium-level features), and 20×20 (deep, low-resolution features, which are beneficial for detecting large targets). Each detection head contains two parallel processing paths with identical structures and non-shared parameters. Each path consists of the following components connected in sequence: the first-level CBS module performs preliminary feature transformation and enhancement on the input features; the second-level CBS module further extracts and integrates high-level semantic features to prepare for the final prediction; and the two-dimensional convolutional layer at the end of the path maps the features extracted by the previous CBS module to the final prediction output space.

[0124] In another embodiment of this application, a target detection method based on a neural network is provided. This method uses a physical guidance channel intelligent dynamic self-pruning neural network as described in any of the foregoing embodiments. The method includes the following steps in sequence:

[0125] S1. Input the image to be detected into the backbone network of the neural network.

[0126] In some embodiments, the input size of the image to be detected is preferably 640×640 pixels and has 3 color channels (RGB); the backbone network performs layer-by-layer feature extraction and abstraction on the input image.

[0127] S2. Configurable multi-scale feature fusion is performed through an open-path-based adjustable multi-scale feature extraction module (CMOC) deployed in the backbone network.

[0128] Specifically, the CMOC module receives the initial feature map extracted from the front end of the backbone network, and performs feature extraction through multiple parallel convolutional pathways (e.g., 3×3, 5×5, 7×7, 9×9) with different kernel sizes. The activation state of each pathway is dynamically controlled using an adjustable parameter set {K1, K2, K3, K4}. The feature maps output by all activated pathways are concatenated along the channel dimension while maintaining spatial consistency, achieving flexible and rich multi-scale feature fusion and enhancing the model's ability to represent targets at different scales.

[0129] S3. The feature channels are dynamically optimized and pruned by the Physical Information Guided Self-Pruning Soft Channel Attention Mechanism (PSCA) module deployed in the backbone network.

[0130] The PSCA module receives the fused multi-scale feature map processed by the CMOC module. First, it calculates the spatial weights of the feature map through soft pooling to generate an initial channel weight vector. Then, it performs a nonlinear transformation on the vector through a multilayer perceptron (MLP) to generate the final channel importance weights. Finally, it weights each channel of the input feature map according to these weights and suppresses or prunes redundant channels with weights below a preset or learnable threshold, thereby preserving the feature information that is key to the detection task and improving model efficiency.

[0131] S4. The multi-scale features extracted and optimized by the backbone network are fused through the neck network.

[0132] The neck network receives feature maps from different layers of the backbone network (typically including high-resolution features in shallow layers and low-resolution, high-semantic features in deep layers), adjusts the feature map size through upsampling, and fuses the shallow and deep features through channel splicing. During this process, at least one CMOC module and / or PSCA module integrated in the neck network can be called again to further extract multi-scale features and perform dynamic channel optimization on the features in the fusion process.

[0133] S5. Output the final target category and location prediction results through the detection head.

[0134] The detection head contains three independent branches, which are respectively connected to feature maps of three different scales (e.g., 80×80, 40×40, 20×20) output by the neck network. Each branch processes the input features through its internal convolutional layer, predicting the class probability and bounding box coordinates of the target at the corresponding scale in parallel. Finally, the prediction results of all branches are integrated to form a complete description of all detected targets in the input image.

[0135] Example

[0136] 1) Neural Network Construction

[0137] In this example, the input resolution of the constructed neural network is set to 640×640×3 (width×height×channels). The network structure is referenced below. Figure 1 As shown, it consists of three parts: the backbone network, the neck network, and the head.

[0138] 1.1 Backbone Network

[0139] The backbone network is responsible for feature extraction, and its specific composition and data flow are as follows:

[0140] Initial downsampling: The input image first undergoes preliminary feature extraction and downsampling through two CBS modules. The first CBS module (k=3, s=2, p=1) downsamples the input from 640×640 to 320×320. The second CBS module (with the same parameters) further downsamples to 160×160.

[0141] Feature extraction and fusion: The feature maps are then processed sequentially through three C3k2 modules (output sizes of 160×160, 80×80, and 40×40) for deep feature extraction.

[0142] Multi-Scale Feature Extraction (CMOC): Next, the feature map is input to the adjustable multi-scale feature extraction module (CMOC) based on an open-path mechanism. This module contains four parallel convolutional branches with kernel sizes (k) of 3, 5, 7, and 9, a stride (s) of 2, and padding (p) of 1, 2, 3, and 4 to ensure consistent output size. In this example, the adjustable parameter set {K1,K2,K3,K4}={1,1,1,1} is set, activating all four branches. Each branch downsamples the input (40×40), resulting in an output size of 20×20 while maintaining the same number of channels. The outputs of the four branches are concatenated along the channel dimension to generate a feature map (20×20) that integrates multiple receptive fields, with four times the number of input channels.

[0143] Channel Dynamic Optimization: The output of the CMOC module is then fed into the Physical Information Guided Self-Pruning Channel Attention (PSCA) module. This module first calculates the spatial weight matrix of the feature map through soft pooling, and then generates the initial channel weight vector through weighted summation.

[0144] The initial weight vector is then input into a three-layer multilayer perceptron (MLP) with the structure: fully connected layer -> SiLU activation -> fully connected layer -> SiLU activation -> fully connected layer) for nonlinear transformation to generate the final channel importance weight vector. y .

[0145] Finally, each channel of the original input feature map is associated with its corresponding channel importance weight vector. y Multiply the values ​​and set a threshold to suppress channels with weights below this threshold (i.e., set the feature value of the channel to zero) to achieve dynamic pruning.

[0146] Deep feature enhancement: The pruned feature map is then further refined and enhanced by a C3k2 module, a spatial pyramid fast pooling module SPPF, and a cross-level pyramid slicing attention module C2PSA. After all operations, the feature map size remains 20×20.

[0147] 1.2 Neck network

[0148] The neck network is responsible for multi-scale feature fusion, and its structure is shown in the attached figure. Figure 1 As shown, the 20×20 feature map output from the end of the backbone network is upsampled to obtain a 40×40 feature map. This upsampled feature map is then concatenated with a shallow feature map of the same 40×40 scale in the backbone network. The concatenated feature map is then fused through one C3k2 module. The fused feature map (40×40) is upsampled again to 80×80 and concatenated with an 80×80 shallow feature map in the backbone network. The concatenated feature map is then fused through another C3k2 module.

[0149] In addition, the neck network in this example also integrates a CMOC module (setting {K1,K2,K3,K4}={1,1,0,0}, i.e. only activating the 3×3 and 5×5 convolutional pathways) and a PSCA module for additional multi-scale extraction and channel optimization of features during the fusion process.

[0150] 1.3 Detection Head

[0151] The detection head section uses three structurally identical branches, which are respectively connected to the feature maps of three scales (80×80, 40×40, and 20×20) output by the neck network (see attached diagram). Figure 1 (As shown). Each branch consists of two CBS modules and one final Conv2d layer, responsible for predicting the target's class probability and bounding box coordinates at the corresponding scale.

[0152] 2) Model Training and Inference

[0153] Training: The network built for this example was trained using the Underwater Target Detection Dataset (DUO). During training, the loss function adopted from the original YOLOv11 loss function, including classification loss, confidence loss, and bounding box regression loss. The optimizer used was AdamW, with an initial learning rate of 1e-3 and a cosine annealing strategy for decreasing the learning rate.

[0154] Inference: After training, the image to be detected is scaled to 640×640×3 and input into the network. The image is processed sequentially through the backbone network and the neck network, and finally the prediction results are output by the three branches of the detection head. After applying non-maximum suppression to the prediction results at the three scales, the final object detection result is obtained, including the class label and bounding box location. A typical object detection visualization is shown below. Figure 5 As shown, the images contain two types of targets: starfish and sea urchins. Starfish are identified using black detection boxes, while sea urchins are identified using white detection boxes. Although the three underwater optical images suffer from color distortion and blurred details due to seawater absorption and scattering effects, the proposed model can still accurately identify both starfish and sea urchins in the images.

[0155] This example achieves flexible fusion of multi-scale features and channel-level dynamic optimization pruning by introducing a configurable CMOC module and a soft-pooling-based CMOC module. Compared with current state-of-the-art models, the metrics are shown in Table 1. The model achieves optimal performance in mAP50, mAP50~95, and average recall, with values ​​of 0.845, 0.658, and 0.764, respectively.

[0156] Table 1 Comparison of Indicators of Different Models

[0157]

[0158] Although the embodiments of this application have been described above in conjunction with the accompanying drawings, this application is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of this application, and these are all within the scope of protection of this application.

Claims

1. A smart dynamic self-pruning neural network for a physical guidance channel, characterized in that, include: The backbone network is used for hierarchical feature extraction from the input image; The neck network, with its input end connected to the output end of the backbone network, is used to receive and fuse shallow and deep features from different depths of the backbone network. The detection head, with its input end connected to the output end of the neck network, is used to predict the target category and location based on the fused multi-scale feature map; The backbone network includes: Multiple front-end feature extraction modules are used to perform preliminary downsampling and feature mapping on the input image; The Configurable Multi-Scale Feature Extraction Module (CMOC) based on the open-path mechanism has its input connected to the output of the front-end feature extraction module. It is used to receive the preliminary feature map and extract and fuse features from different receptive fields through multiple parallel convolutional paths that are independently controlled by adjustable parameters. The Physical Information Guided Self-Pruning Channel Attention Module (PSCA) has its input connected to the output of the CMOC module. It is used to perform soft pooling on the fused multi-scale feature map to calculate spatial weights and generate channel weights through a multilayer perceptron. Based on this, the feature channels are dynamically pruned to retain key features. Multiple back-end feature extraction and fusion modules, whose inputs are connected to the output of the PSCA module, are used to further process the pruned and optimized features and output multi-scale deep features to the neck network.

2. The intelligent dynamic self-pruning neural network for physical guidance channels according to claim 1, characterized in that, The CMOC module includes: Multiple parallel convolutional paths, each corresponding to a convolutional kernel of a specific size, are interconnected at their input ends and jointly receive the input feature map; The channel stitching layer, whose input is connected to the output of all activated parallel convolutional paths, is used to stitch the feature maps output by each path along the channel dimension. Each convolutional path is equipped with a switch unit that can be controlled by external parameters. When the switch unit is in the first state, the path is activated and its output is transmitted to the channel splicing layer. When the switch unit is in the second state, the path is disabled.

3. The intelligent dynamic self-pruning neural network for physical guidance channels according to claim 2, characterized in that, The multiple parallel convolutional pathways include convolutional layers with kernel sizes of 3×3, 5×5, 7×7 and 9×9, respectively.

4. The intelligent dynamic self-pruning neural network for physical guidance channels according to claim 1, characterized in that, The PSCA module includes: The soft pooling unit, whose input is connected to the output of the CMOC module, is used to calculate the probability weight of each spatial location in the feature map through the Softmax function, and generate the initial channel weight vector by weighted summation. A multilayer perceptron (MLP) has its input connected to the output of the soft pooling unit, and is used to perform a nonlinear transformation on the initial channel weight vector to generate the final channel importance weight vector. The channel weighting and pruning unit has two input terminals. Its first input terminal receives the output feature map from the CMOC module, and its second input terminal is connected to the output terminal of the multilayer perceptron (MLP) to receive the channel importance weight vector, which is used to multiply the feature map of each channel with the corresponding weight and suppress redundant channels with weights below a threshold.

5. The intelligent dynamic self-pruning neural network for physical guidance channels according to claim 4, characterized in that, The soft pooling unit calculates the first channel weight vector in the initial channel weight vector. m The weight of each channel x m for: in, B m Indicates the first m Spatial weight matrix for each channel, each element b i,j The spatial weights are calculated using the Softmax function and represent the positions in the feature map. i, j The importance of ); A m Indicates the first m The original eigenvalue matrix of each channel, each element a i,j It is the first m Each channel is located at ( i, j The eigenvalue at () location; R This represents the spatial dimension of the entire feature map.

6. The intelligent dynamic self-pruning neural network for physical guidance channels according to claim 4, characterized in that, The multilayer perceptron (MLP) consists of three fully connected layers, with the SiLU function as the activation function.

7. The intelligent dynamic self-pruning neural network for physical guidance channels according to claim 1, characterized in that, The backend feature extraction module in the backbone network includes cascaded C3k2 modules; The backbone network is also connected to a Spatial Pyramid Fast Pooling (SPPF) module and a Cross-Level Pyramid Slice Attention (C2PSA) module at its end. The input of the SPPF module is connected to the output of the back-end feature extraction module, and the input of the C2PSA module is connected to the output of the SPPF module. The output of the C2PSA module is connected to the neck network.

8. The intelligent dynamic self-pruning neural network for physical guidance channels according to claim 1, characterized in that, The neck network increases the size of the deep feature map through upsampling and fuses it with the feature map from the shallow layer of the backbone network through channel splicing. The neck network integrates at least one CMOC module and / or at least one PSCA module for multi-scale extraction and dynamic optimization of features during the fusion process.

9. The intelligent dynamic self-pruning neural network for physical guidance channels according to claim 1, characterized in that, The detection head includes multiple branches, which are respectively connected to feature maps of different scales output by the neck network, for target detection at multiple resolutions.

10. A target detection method based on a neural network, characterized in that, Using the intelligent dynamic self-pruning neural network for the physical guidance channel according to any one of claims 1 to 9, the method includes: The image to be detected is input into the backbone network of the neural network; Configurable multi-scale feature fusion is performed through the CMOC module in the backbone network; The feature channels are dynamically pruned using the PSCA module in the backbone network. The neck network is fused with multi-scale features extracted from the backbone network; The detection head outputs the final target category and location prediction results.

Images