Parameter efficient fine-tuning target detection method and system based on weight matrix convolution

Abstract

The application provides a parameter efficient fine-tuning target detection method based on weight matrix convolution, which comprises the following steps: loading a pre-trained Transformer target detection backbone network, only initializing a small two-dimensional learnable convolution kernel, and assigning values to the convolution kernel in a He initialization manner; performing zero padding of the same size on the original convolution kernel; performing two-dimensional spatial convolution operation on the padded convolution kernel and the original weight matrix to generate a weight increment of the same size; introducing a fixed scaling coefficient to constrain the increment amplitude, obtaining enhanced features after spatial convolution weight adaptation; inputting the enhanced features into a detection head to perform target classification and bounding box regression; based on the reverse gradient of the detection loss, only updating the convolution kernel parameters, freezing the backbone weights throughout, and iteratively updating until the model converges; after the training is completed, the incremental weights are merged into the pre-trained backbone to eliminate the additional calculation branch of inference; inputting an image to be detected, and outputting the target category and coordinate detection result. Compared with the prior art, the application can efficiently fine-tune the pre-trained Transformer backbone with very few trainable parameters, and improve the target detection accuracy in various complex scenarios.

Description

Technical Field

[0001] This invention relates to computer vision and deep learning model fine-tuning techniques, and particularly to a method for efficient parameter fine-tuning of target detection based on weight matrix convolution. Background Technology

[0002] With the rapid development of deep learning technology, pre-trained Transformer models, with their powerful general representation capabilities, are widely used in computer vision tasks such as image object detection. To reduce the computational and storage consumption caused by full fine-tuning, parameter efficient fine-tuning (PEFT) technology has become the mainstream solution for transferring pre-trained models to practical applications, among which low-rank fine-tuning methods, represented by LoRA, are the most widely used.

[0003] Existing low-rank fine-tuning methods construct weight update residuals by decomposing the weights into two sets of low-dimensional matrices. Only a small number of parameters participate in training, effectively controlling fine-tuning costs and achieving stable results in tasks such as image classification and basic object detection. However, these methods restrict all weight update operations to a low-dimensional subspace, failing to adequately mine the two-dimensional local spatial correlation features inherent in the pre-trained weight matrix itself. Object detection tasks are characterized by multi-scale targets, complex background occlusion, and variable target shapes, demanding higher representational richness for weight updates. Traditional low-rank constraints can limit the model's adaptability to some extent. Furthermore, some existing schemes embed convolutional structures within the low-rank fine-tuning framework to achieve feature adaptation, but their convolutional operations still rely on the compressed low-dimensional space, failing to fully preserve the spatial structure information of the original weight matrix, and easily leading to feature loss during fine-tuning.

[0004] In summary, there is an urgent need to propose a parameter-efficient fine-tuning target detection method based on weight matrix convolution to solve the above problems. Summary of the Invention

[0005] In view of this, the main objective of this invention is to provide a parameter-efficient fine-tuning target detection method based on weight matrix convolution, which can solve the problems of insufficient feature adaptation capability when existing LoRA-type low-rank fine-tuning schemes constrain weight updates to low-dimensional subspaces, compress the original two-dimensional spatial association information of pre-trained weights, and face multi-scale and occluded complex target detection scenarios.

[0006] To achieve the above objectives, according to a first aspect of the present invention, a parameter-efficient fine-tuning target detection method based on weight matrix convolution is provided, the method comprising: The first step is to load the pre-trained Transformer object detection backbone network, freeze all the original weight matrices of the network, and initialize only small two-dimensional learnable convolutional kernels as parameters to be trained; the He initialization method is used to assign values ​​to the convolutional kernels. The second step is to perform zero padding of the same size on the original convolution kernel to obtain a padded convolution kernel; The third step is to perform a two-dimensional spatial convolution operation between the filled convolution kernel and the original weight matrix to generate weight increments of the same size. The fourth step involves introducing a fixed scaling factor to constrain the increment magnitude, and obtaining the enhanced features after spatial convolution weight adaptation based on the original weight matrix and weight increment. The fifth step involves inputting the enhanced features into the detection head to perform target classification and bounding box regression; based on the inverse gradient of the detection loss, only the convolution kernel parameters are updated, while the backbone weights are frozen throughout the process, and the iteration continues until the model converges. The sixth step is to merge the incremental weights into the pre-training backbone after training is completed, eliminating the extra computational branches of inference; input the image to be detected, and output the target category and coordinate detection results.

[0007] The first step includes: The steps for obtaining the training set, validation set, and test set are as follows: Load the general image object detection dataset, divide all samples into training set, validation set, and test set according to a preset ratio, use the training set for updating model parameters, use the validation set for monitoring the model convergence status, and use the test set for final performance evaluation. The steps to obtain the original weight matrix are as follows: Load the pre-trained Transformer backbone network, set all weights of the backbone network to a frozen state, and read all the original weight matrices. Initial size is Small learnable convolutional kernels The He initialization method is used to assign values ​​to the convolution kernel parameters, where... , These represent the number of rows and columns of the weight matrix, respectively, which are the output and input dimensions of the current layer. Indicates the kernel size. .

[0008] Furthermore, the convolution kernel size The value range is 13 to 19.

[0009] The second step includes: to The original convolutional kernel is zero-paddinged to the same size to generate a padded convolutional kernel with the same dimensions as the weight matrix: ,in , These represent the row index and column index of the weight matrix, respectively. Indicates learnable convolutional kernels In the Line number The elements of the column.

[0010] The third step includes: filling the convolutional kernel Compared with the original weight matrix Perform a 2D sliding convolution operation to generate weight increments that perfectly match the dimensions. ,in This represents a two-dimensional spatial convolution operator.

[0011] The fourth step includes: setting a fixed scaling factor. Constrain the magnitude of the weight increment, based on the original weight matrix. and weight increment The enhanced features obtained after spatial convolution weight adaptation are obtained. ,in This represents the input image features of the Transformer layer, i.e., the image tensor obtained after preprocessing by the backbone network.

[0012] Furthermore, the fixed scaling factor The value range is 0.05 to 0.2.

[0013] The fifth step includes: feeding enhanced features into the detection head module; the detection head performs target category classification and target bounding box regression; calculating the classification loss and bounding box regression loss respectively; and weighted summing them to obtain the overall detection loss; during the backpropagation of loss, only the parameters of small convolutional kernels are updated, while all weights of the backbone network remain unchanged, thereby significantly reducing the number of parameters that need to be trained and reducing GPU memory usage; the training set is used for model parameter updates: samples in the training set are read cyclically according to the set batch size; the learning rate is warmed up in the first 3 rounds, and then linear decay is performed; the validation set is used to monitor the convergence status: after each round of complete training, target detection-related indicators are calculated on the validation set, and early stopping is triggered accordingly; the batch iterative training is repeated until the validation set detection indicators no longer improve, the model is determined to have converged, and the training process ends.

[0014] The sixth step includes: using a test set for final performance evaluation; after training, the weight increments generated by convolution are fused into the original weights of the pre-trained backbone, and the increments are directly merged into the backbone weight parameters. No additional computational branches are retained during the inference stage, and the computational cost of inference is consistent with that of the original pre-trained model; the image to be detected is input into the model, and after the image is processed by the backbone network for feature extraction, the target category, bounding box coordinates, and confidence score are directly output to complete the image target detection task.

[0015] According to another aspect of the present invention, a parameter-efficient fine-tuning target detection system based on weight matrix convolution is provided, the system comprising: The model parameter initialization module is used to load the pre-trained Transformer object detection backbone network, freeze all the original weight matrices of the network, and initialize only small two-dimensional learnable convolutional kernels as parameters to be trained; the He initialization method is used to assign values ​​to the convolutional kernels. The kernel padding module is used to perform zero padding of the same size on the original convolution kernel to obtain a padded convolution kernel; The weight increment acquisition module is used to perform a two-dimensional spatial convolution operation between the filled convolution kernel and the original weight matrix to generate weight increments of the same size. The enhanced feature acquisition module is used to introduce a fixed scaling factor to constrain the incremental magnitude and obtain the enhanced features after spatial convolution weight adaptation based on the original weight matrix and weight increment. The model training module is used to input enhanced features into the detection head and perform target classification and bounding box regression; based on the back gradient of the detection loss, only the convolution kernel parameters are updated, the backbone weights are frozen throughout the process, and the loop iterates until the model converges; The object detection and result output module is used to merge incremental weights into the pre-training backbone after training is completed, eliminating extra computational branches for inference; it takes an image to be detected as input and outputs the object category and coordinate detection results.

[0016] Compared with existing target detection technologies, the advantages of the parameter-efficient fine-tuning target detection method based on weight matrix convolution of the present invention are as follows: it abandons the low-rank matrix decomposition operation, directly generates weight correction values ​​through lightweight convolution in the complete original parameter space, fully preserves the local structural features inside the weights, achieves efficient fine-tuning of the pre-trained Transformer backbone with a very small number of trainable parameters, and improves the target detection accuracy in various complex scenarios. Attached Figure Description

[0017] Figure 1 A flowchart of a parameter-efficient fine-tuning target detection method based on weight matrix convolution according to the present invention is shown.

[0018] Figure 2 A framework diagram of a parameter-efficient fine-tuning target detection system based on weight matrix convolution according to the present invention is shown. Detailed Implementation

[0019] To enable those skilled in the art to further understand the structure, features and other objectives of the present invention, the following detailed description is provided in conjunction with the accompanying preferred embodiments. The preferred embodiments described are only for illustrating the technical solutions of the present invention and are not intended to limit the present invention.

[0020] Figure 1 A flowchart of the parameter-efficient fine-tuning target detection method based on weight matrix convolution according to the present invention is given. Figure 1As shown, the parameter-efficient fine-tuning target detection method based on weight matrix convolution according to the present invention includes: The first step, S1, loads the pre-trained Transformer object detection backbone network, freezes all the original weight matrices of the network, and initializes only small two-dimensional learnable convolutional kernels as parameters to be trained; the He initialization method is used to assign values ​​to the convolutional kernels. The second step, S2, involves performing zero padding of the same size on the original convolutional kernel to obtain a padded convolutional kernel. The third step, S3, involves performing a two-dimensional spatial convolution operation between the filled convolution kernel and the original weight matrix to generate weight increments of the same size. In the fourth step S4, a fixed scaling factor is introduced to constrain the increment magnitude. Based on the original weight matrix and the weight increment, the enhanced features after spatial convolution weight adaptation are obtained. In step S5, the enhanced features are input into the detection head to perform target classification and bounding box regression. Based on the back gradient of the detection loss, only the convolution kernel parameters are updated, while the backbone weights are frozen throughout the process. The iteration continues until the model converges. In step S6, after training is complete, the incremental weights are merged into the pre-training backbone to eliminate the extra computational branches of inference; the input image to be detected is output as the target category and coordinate detection results.

[0021] The first step S1 includes: Step S11: Load the general image object detection dataset and divide all samples into training set, validation set and test set according to the preset ratio. The training set is used for model parameter update, the validation set is used to monitor the model convergence status, and the test set is used for final performance evaluation. Step S12, obtaining the original weight matrix, involves loading the pre-trained Transformer backbone network, freezing all weights in the backbone network, and reading all the original weight matrices. Initial size is Small learnable convolutional kernels The He initialization method is used to assign values ​​to the convolution kernel parameters, where... , These represent the number of rows and columns of the weight matrix, respectively, which are the output and input dimensions of the current layer. Indicates the kernel size. .

[0022] Furthermore, the convolution kernel size The value range is 13 to 19. For example, the convolution kernel size... The answer is 17.

[0023] The He initialization method is a weight initialization method specifically designed for ReLU and its variants (such as Leaky ReLU) activation functions. For example, it is implemented according to the method in the literature "Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification. Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. IEEE International Conference on ComputerVision (ICCV) 2015".

[0024] The second step S2 includes: ... The original convolutional kernel is zero-paddinged to the same size to generate a padded convolutional kernel with the same dimensions as the weight matrix: ,in , These represent the row index and column index of the weight matrix, respectively. Indicates learnable convolutional kernels In the Line number The elements of the column.

[0025] In the second step S2, the smaller learnable convolutional kernel is mapped to the same dimension as the backbone weight matrix, ensuring that the dimensions of subsequent convolution operations can match.

[0026] The third step includes: filling the convolutional kernel Compared with the original weight matrix Perform a 2D sliding convolution operation to generate weight increments that perfectly match the dimensions. ,in This represents a two-dimensional spatial convolution operator.

[0027] The It iterates through the pre-trained weight matrix window by window, performing weighted aggregation within each sliding window to uncover local correlation features within the weight matrix and output structured weight corrections. This operation performs convolution directly in the model's original parameter space, without needing to decompose or transform the pre-trained weights, thus fully preserving the prior knowledge already learned by the pre-trained model.

[0028] Furthermore, the fourth step S4 includes: setting a fixed scaling factor. Constrain the magnitude of the weight increment, based on the original weight matrix. and weight increment The enhanced features are obtained after spatial convolution weight adaptation. ,in This represents the input image features of the Transformer layer, i.e., the image tensor obtained after preprocessing by the backbone network.

[0029] Furthermore, the fixed scaling factor The value range is 0.05 to 0.2. For example, Choose 0.15.

[0030] The enhanced features It integrates the output of the original pre-trained model with the correction information brought by parameter space convolution, and then feeds it into the downstream detection head to complete the target detection task.

[0031] Furthermore, the fifth step S5 includes: enhancing the features The data is fed into the detection head module, which performs target category classification and target bounding box regression. The classification loss and bounding box regression loss are calculated separately and then weighted and summed to obtain the overall detection loss. During backpropagation of the loss, only the parameters of small convolutional kernels are updated, while all weights in the backbone network remain constant. This significantly reduces the number of parameters required for training and minimizes GPU memory usage. The training set is used for model parameter updates: samples from the training set are read cyclically according to the set batch size. The first three rounds perform a learning rate warm-up, followed by linear decay. The validation set is used to monitor convergence: after each complete training round, target detection metrics are calculated on the validation set, and early stopping is triggered accordingly. Batch iterations are repeated until the validation set detection metrics no longer improve, at which point the model is considered converged, and the training process ends.

[0032] The sixth step S6 includes: using a test set for final performance evaluation; after training, the weight increments generated by convolution are fused into the original weights of the pre-trained backbone, and the increments are directly merged into the backbone weight parameters. No additional computation branches are retained during the inference stage, and the computational load of inference is consistent with that of the original pre-trained model; the image to be detected is input into the model, and after the image is processed by the backbone network for feature extraction, the target category, bounding box coordinates, and confidence score are directly output to complete the image target detection task.

[0033] Figure 2 A framework diagram of a parameter-efficient fine-tuning target detection system based on weight matrix convolution according to the present invention is given. Figure 2 As shown, the parameter-efficient fine-tuning target detection system based on weight matrix convolution according to the present invention includes: Model parameter initialization module 1 is used to load the pre-trained Transformer object detection backbone network, freeze all the original weight matrices of the network, and only initialize small two-dimensional learnable convolutional kernels as parameters to be trained; the He initialization method is used to assign values ​​to the convolutional kernels. Convolution kernel padding module 2 is used to perform zero padding of the same size on the original convolution kernel to obtain a padded convolution kernel; The weight increment acquisition module 3 is used to perform a two-dimensional spatial convolution operation between the filled convolution kernel and the original weight matrix to generate a weight increment of the same size. The enhanced feature acquisition module 4 is used to introduce a fixed scaling factor to constrain the incremental magnitude and to obtain the enhanced features after spatial convolution weight adaptation based on the original weight matrix and weight increment. Model training module 5 is used to input enhanced features into the detection head and perform target classification and bounding box regression; based on the back gradient of the detection loss, only the convolution kernel parameters are updated, the backbone weights are frozen throughout, and the process is iterated until the model converges. The target detection and result output module 6 is used to merge the incremental weights into the pre-training backbone after training is completed, eliminating the extra calculation branches of inference; it takes the image to be detected as input and outputs the target category and coordinate detection results.

[0034] The model parameter initialization module 1 includes: The training set, validation set and test set acquisition module 11 is used to load a general image object detection dataset and divide all samples into training set, validation set and test set according to a preset ratio. The training set is used for model parameter update, the validation set is used to monitor the model convergence status and the test set is used for final performance evaluation. The original weight matrix acquisition module 12 is used to load the pre-trained Transformer backbone network, set all weights of the backbone network to a frozen state, and read all the original weight matrices. Initial size is Small learnable convolutional kernels The He initialization method is used to assign values ​​to the convolution kernel parameters, where... , These represent the number of rows and columns of the weight matrix, respectively, which are the output and input dimensions of the current layer. Indicates the kernel size. .

[0035] Furthermore, the convolution kernel size The value range is 13 to 19.

[0036] The convolution kernel filling module 2 includes: a function for filling kernels... The original convolutional kernel is zero-paddinged to the same size to generate a padded convolutional kernel with the same dimensions as the weight matrix: ,in , These represent the row index and column index of the weight matrix, respectively. Indicates learnable convolutional kernels In the Line number The elements of the column.

[0037] The weight increment acquisition module 3 includes: a module for obtaining the filled convolutional kernel. Compared with the original weight matrix Perform a 2D sliding convolution operation to generate weight increments that perfectly match the dimensions. ,in This represents a two-dimensional spatial convolution operator.

[0038] The enhanced feature acquisition module 4 includes: setting a fixed scaling factor. Constrain the magnitude of the weight increment, based on the original weight matrix. and weight increment The enhanced features obtained after spatial convolution weight adaptation are obtained. ,in This represents the input image features of the Transformer layer, i.e., the image tensor obtained after preprocessing by the backbone network.

[0039] Furthermore, the fixed scaling factor The value range is 0.05 to 0.2.

[0040] The model training module 5 includes: a module for training enhanced features. The data is fed into the detection head module, which performs target category classification and target bounding box regression. The classification loss and bounding box regression loss are calculated separately and then weighted and summed to obtain the overall detection loss. During backpropagation of the loss, only the parameters of small convolutional kernels are updated, while all weights in the backbone network remain constant. This significantly reduces the number of parameters required for training and minimizes GPU memory usage. The training set is used for model parameter updates: samples from the training set are read cyclically according to the set batch size. The first three rounds perform a learning rate warm-up, followed by linear decay. The validation set is used to monitor convergence: after each complete training round, target detection metrics are calculated on the validation set, and early stopping is triggered accordingly. Batch iterations are repeated until the validation set detection metrics no longer improve, at which point the model is considered converged, and the training process ends.

[0041] The target detection and result output module 6 includes: a test set for final performance evaluation; after training, the weight increments generated by convolution are fused into the original weights of the pre-trained backbone, and the increments are directly merged into the backbone weight parameters. No additional computation branches are retained during the inference stage, and the inference computation is consistent with the original pre-trained model; the image to be detected is input into the model, and after the image is processed by the backbone network feature extraction, the target category, bounding box coordinates and confidence score are directly output to complete the image target detection task.

[0042] Compared with existing target detection technologies, the advantages of the parameter-efficient fine-tuning target detection method based on weight matrix convolution of the present invention are as follows: it abandons the low-rank matrix decomposition operation, directly generates weight correction values ​​through lightweight convolution in the complete original parameter space, fully preserves the local structural features inside the weights, achieves efficient fine-tuning of the pre-trained Transformer backbone with a very small number of trainable parameters, and improves the target detection accuracy in various complex scenarios.

[0043] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It should be understood that the present invention is not limited to the implementations described herein. These implementations are described to help those skilled in the art practice the present invention. Any person skilled in the art can easily make further improvements and modifications without departing from the spirit and scope of the present invention. Therefore, the present invention is only limited by the content and scope of the claims, and is intended to cover all alternatives and equivalents included within the spirit and scope of the present invention as defined by the appended claims.

Claims

1. A parameter-efficient fine-tuning target detection method based on weight matrix convolution, characterized in that, The method includes: The first step is to load the pre-trained Transformer object detection backbone network, freeze all the original weight matrices of the network, and initialize only small two-dimensional learnable convolutional kernels as parameters to be trained; the He initialization method is used to assign values ​​to the convolutional kernels. The second step is to perform zero padding of the same size on the original convolution kernel to obtain a padded convolution kernel; The third step is to perform a two-dimensional spatial convolution operation between the filling convolution kernel and the original weight matrix to generate weight increments of the same size. The fourth step involves introducing a fixed scaling factor to constrain the increment magnitude, and obtaining the enhanced features after spatial convolution weight adaptation based on the original weight matrix and weight increment. The fifth step involves inputting the enhanced features into the detection head to perform target classification and bounding box regression; based on the inverse gradient of the detection loss, only the convolution kernel parameters are updated, while the backbone weights are frozen throughout the process, and the iteration continues until the model converges. The sixth step is to merge the incremental weights into the pre-training backbone after training is completed, eliminating the extra computational branches of inference; input the image to be detected, and output the target category and coordinate detection results.

2. The method as described in claim 1, characterized in that, The first step includes: The steps for obtaining the training set, validation set, and test set are as follows: Load the general image object detection dataset, divide all samples into training set, validation set, and test set according to the preset ratio. The training set is used for updating model parameters, the validation set is used to monitor the model convergence status, and the test set is used for final performance evaluation. The steps to obtain the original weight matrix are as follows: Load the pre-trained Transformer backbone network, set all weights of the backbone network to a frozen state, and read all the original weight matrices. Initial size is Small learnable convolutional kernels The He initialization method is used to assign values ​​to the convolution kernel parameters, where , These represent the number of rows and columns of the weight matrix, respectively, which are the output and input dimensions of the current layer. Indicates the kernel size. .

3. The method as described in claim 2, characterized in that, The second step includes: to The original convolutional kernel is zero-paddinged to the same size to generate a padded convolutional kernel with the same dimensions as the weight matrix: ,in , These represent the row index and column index of the weight matrix, respectively. Indicates learnable convolutional kernels In the Line 1 The elements of the column.

4. The method as described in claim 3, characterized in that, The third step includes: filling the convolutional kernel Compared with the original weight matrix Perform a 2D sliding convolution operation to generate weight increments that perfectly match the dimensions. ,in This represents a two-dimensional spatial convolution operator.

5. The method as described in claim 4, characterized in that, The fourth step includes: setting a fixed scaling factor. Constrain the magnitude of the weight increment, based on the original weight matrix. and weight increment The enhanced features obtained after spatial convolution weight adaptation are obtained. ,in This represents the input image features of the Transformer layer, i.e., the image tensor obtained after preprocessing by the backbone network; Furthermore, the fixed scaling factor The value range is 0.05 to 0.

2.

6. The method as described in claim 1, characterized in that, The fifth step includes: feeding enhanced features into the detection head module; the detection head performs target category classification and target bounding box regression; calculating the classification loss and bounding box regression loss respectively; and weighted summing them to obtain the overall detection loss; during the backpropagation of loss, only the parameters of small convolutional kernels are updated, while all weights of the backbone network remain unchanged, thereby significantly reducing the number of parameters that need to be trained and reducing GPU memory usage; the training set is used for model parameter updates: samples in the training set are read cyclically according to the set batch size; the learning rate is warmed up in the first 3 rounds, and then linear decay is performed; the validation set is used to monitor the convergence status: after each round of complete training, target detection-related indicators are calculated on the validation set, and early stopping is triggered accordingly; the batch iterative training is repeated until the validation set detection indicators no longer improve, the model is determined to have converged, and the training process ends.

7. The method as described in claim 1, characterized in that, The sixth step includes: using a test set for final performance evaluation; after training, the weight increments generated by convolution are fused into the original weights of the pre-trained backbone, and the increments are directly merged into the backbone weight parameters. No additional computational branches are retained during the inference stage, and the computational cost of inference is consistent with that of the original pre-trained model; the image to be detected is input into the model, and after the image is processed by the backbone network for feature extraction, the target category, bounding box coordinates, and confidence score are directly output to complete the image target detection task.

8. The method as described in claim 2, characterized in that, The convolution kernel size The value range is 13 to 19.

9. A parameter-efficient fine-tuning target detection system based on weight matrix convolution, characterized in that, The system includes: The model parameter initialization module is used to load the pre-trained Transformer object detection backbone network, freeze all the original weight matrices of the network, and initialize only small two-dimensional learnable convolutional kernels as parameters to be trained; the He initialization method is used to assign values ​​to the convolutional kernels. The kernel padding module is used to perform zero padding of the same size on the original convolution kernel to obtain a padded convolution kernel; The weight increment acquisition module is used to perform a two-dimensional spatial convolution operation between the filled convolution kernel and the original weight matrix to generate a weight increment of the same size. The enhanced feature acquisition module is used to introduce a fixed scaling factor to constrain the incremental magnitude and obtain the enhanced features after spatial convolution weight adaptation based on the original weight matrix and weight increment. The model training module is used to input enhanced features into the detection head and perform target classification and bounding box regression; based on the back gradient of the detection loss, only the convolution kernel parameters are updated, the backbone weights are frozen throughout the process, and the loop iterates until the model converges; The object detection and result output module is used to merge incremental weights into the pre-training backbone after training is completed, eliminating extra computational branches for inference; it takes an image to be detected as input and outputs the object category and coordinate detection results.

Images