Target identification method based on deep learning in complex environment

Abstract

The invention discloses a deep learning-based target recognition method under a complex environment condition, relates to a target recognition method, and aims at improving the recognition method under the water mist and low-illumination shape conditions based on the problems of poor recognition precision and low robustness of traditional target recognition under the complex environment condition. On the basis of image enhancement, the robustness is improved, and the method can be used for small target recognition. Based on the improvement, when an image enters an algorithm structure, an advanced gating mechanism structure module can differentiate an image processing module, a global attention mechanism is added, and final enhanced output is obtained through output. As input of a target detection model, a subsequent target identification task is carried out. The method is suitable for related target recognition in a complex environment, can also be used for target recognition of small targets, and has relatively high robustness and recognition precision.

Description

Technical Field

[0001] This invention relates to a target recognition method, and more particularly to a deep learning-based target recognition method for complex environments. Background Technology

[0002] With the continuous advancement of deep learning, the requirements for accuracy, speed, and robustness in target recognition are increasing. Traditional target recognition methods cannot meet the accuracy needs of current practical production in complex environments. However, with the improvement of target detection algorithms, especially the Convolutional Neural Network (CNN) structure in deep learning models, which can automatically learn target features from large amounts of data without manual feature design and extraction, deep learning models trained on large-scale data have strong generalization capabilities and can demonstrate good performance on different datasets. Compared with traditional target recognition methods, deep learning models have significantly improved the accuracy of target recognition. Deep learning models trained on large amounts of data can learn the complex feature patterns of various targets, thus accurately identifying different types of targets. This represents a major breakthrough in the field of target recognition. It can automatically learn features in images, greatly improving the accuracy and efficiency of target recognition.

[0003] In practical applications, changes in lighting conditions, such as strong light, shadow, and backlight, as well as adverse weather conditions, can lead to a decrease in the accuracy of target recognition. In some cases, the model is also prone to overfitting, resulting in insufficient generalization ability for some data and difficulty in accurately identifying targets. Summary of the Invention

[0004] The purpose of this invention is to provide a target recognition method based on deep learning in complex environments. This invention effectively reduces image noise and preserves low-level features through a gating mechanism module; improves image quality by using a global module for color reconstruction and gamma correction; and introduces a loss module; thereby improving the robustness and accuracy of target recognition in complex environments.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A deep learning-based target recognition method for complex environments is disclosed. The method employs a general enhancement module comprising a gating mechanism module, a global module, and a loss module. The gating mechanism module takes the original image as input and uses the gating mechanism to concurrently perform relative weighting on multiple differentiable image processing modules to enhance the image under adverse environmental conditions, resulting in an output image. The global module's input consists of the feature image obtained above and the original image, and its output is the final enhanced output. The loss module takes the enhanced output and the original image as input and is connected after the gating mechanism module. The model's weights are continuously optimized through training. Finally, the enhanced output is used as the input to the target detection model for subsequent target recognition tasks.

[0007] The aforementioned deep learning-based target recognition method in complex environments includes a gating mechanism module further improved by multiple GDIPs through a multi-stage guided process for progressive image enhancement. To balance accuracy and speed while maintaining speed, a variant of the GDIP is proposed, eliminating the need for GDIP-based image enhancement during inference, thus achieving higher throughput and feasible real-world deployment. Multi-level progressive image enhancement is achieved by combining GDIP blocks with each layer of the visual encoder.

[0008] The aforementioned deep learning-based target recognition method in complex environments includes a global module whose input consists of a key and a value, where the key refers to the original image and the value refers to the feature image. The global branch consists of a first attention module and a second attention module, wherein the attention modules are linearly connected and employ the ACmis self-attention mechanism, which combines convolution and self-attention aggregation. The 1x1 convolution is shared between the two modules, aiming to reduce computational overhead and integrate lightweight aggregation operations.

[0009] The aforementioned deep learning-based target recognition method in complex environments comprises a loss module consisting of three parts: smoothing loss, structure-aware smoothing loss, and consistency loss. First, the gradient intensity T of the original image is calculated using the gradient intensity calculation formula. s and noise reduction features T D .

[0010] The aforementioned deep learning-based target recognition method in complex environments uses a smoothing loss formula L. smooth =||T D ||, where ‖·‖ refers to the L2 norm.

[0011] The aforementioned deep learning-based target recognition method in complex environments includes a structure-aware smoothing loss formula as follows: Here, λ is a hyperparameter representing the structure-sensing intensity. This refers to multiplying by element position.

[0012] The advantages and effects of this invention are:

[0013] 1. This invention addresses the shortcomings of traditional target recognition methods in complex environments, such as poor accuracy and low robustness, by improving recognition methods under conditions of water mist and low light. Building upon image enhancement, robustness is further improved for small target recognition. Based on these improvements, when an image enters the algorithm structure, an advanced gating mechanism module differentiates the image processing module, adds a global attention mechanism, and outputs the final enhanced output. This enhanced output serves as input to the target detection model for subsequent target recognition tasks. The gating mechanism module effectively reduces image noise while preserving low-level features; the global module performs color reconstruction and gamma correction to improve image quality; a loss module is introduced to balance image smoothness and texture feature preservation; and the module can be added as an add-on to existing target detection models to increase robustness in various environments.

[0014] 2. This invention is applicable to target recognition in complex environments, and can also be used for target recognition of small targets, with strong robustness and recognition accuracy. Attached Figure Description

[0015] Figure 1 This is an overall structural diagram of the gating mechanism module of the present invention;

[0016] Figure 2 This is a partial structural diagram of the global module of the present invention;

[0017] Figure 3 This is a diagram of the overall framework of the model of this invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the embodiments shown in the accompanying drawings.

[0019] The general enhancement module employed in this invention includes a gating mechanism module, a global module, and a loss module. The gating mechanism module takes the original image as input and employs a novel gating mechanism that allows multiple differentiable image processing modules to concurrently perform relative weighting to enhance images under adverse environmental conditions and output the final enhanced image. The global module's input consists of the feature image obtained above and the original image, and its output is the final enhanced output. The loss module takes the enhanced output and the original image as input. The loss module is connected after the gating mechanism module; its weights are continuously optimized through training. Finally, the obtained enhanced output serves as the input to the object detection model for subsequent object recognition tasks.

[0020] The gating mechanism module is further improved by multiple GDIPs through a multi-stage bootstrapping process for progressive image enhancement. For example... Figure 1 The diagram shows the overall structure of the gating mechanism module. While maintaining speed, a variant of GDIP is proposed to balance accuracy and speed. This eliminates the need for GDIP-based image augmentation during inference, thereby achieving higher throughput and feasible real-world deployment. Multi-level progressive image augmentation is achieved by integrating the GDIP block with each layer of the visual encoder. This design allows GDIP to leverage feature representations from different levels to guide various image processing operations. In this way, the system can process different features in the image more finely, such as edges or textures, thereby improving object detection performance under diverse environmental conditions.

[0021] The input to the global module consists of a key and a value, where the key refers to the original image and the value refers to the feature image. The global branch consists of a first attention module and a second attention module, which are connected linearly. The attention modules employ the ACmis self-attention mechanism, such as... Figure 2 As shown, it combines convolution and self-attention aggregation, with a 1x1 convolution shared between the two, aiming to reduce computational overhead and integrate lightweight aggregation operations. The keys and values, after being embedded in the residual positions, are used as the key and value inputs to the first attention module, resulting in a 1×10 shape first sequence. This first sequence is then used as the key and value input to the second attention module, and its output is added to the first sequence, resulting in a 1×10 shape second sequence. This second sequence is then split and reshaped to obtain a 3×3 color convolution kernel and a 1×1 gamma value. The color convolution kernel is convolved with the denoising features to obtain color reconstruction features. These color features are finally gamma-corrected based on the gamma value, resulting in an enhanced output.

[0022] The loss module consists of three parts: smoothing loss, structure-aware smoothing loss, and consistency loss. First, the gradient intensity T of the original image is calculated according to the gradient intensity calculation formula. s and the noise reduction feature T D The formula for smoothing loss is L. smooth =||T D ||, where ‖·‖ refers to the L2 norm; the formula for the structure-aware smoothing loss is: Here, λ is a hyperparameter representing the structure-sensing intensity. This refers to multiplying by element position; the formula for the consistency loss is: in, This refers to subtracting elements by their positions; finally, the total loss of the loss module is expressed as: L = L smooth +L sp+L c The location embedding module adopts a sequential structure, with each layer from front to back consisting of a 3×3 convolution, a batch normalization layer, a GELU activation function layer, another 3×3 convolution, and a batch normalization layer.

[0023] The formula for calculating gradient strength is: Where I represents the input, This indicates the calculation of the gradient in the horizontal and vertical directions of the input. Figure 3 This is a diagram of the overall framework of the model of this invention.

Claims

1. A target recognition method in complex environments based on deep learning, characterized in that: The general enhancement module adopted by the method includes a gating mechanism module, a global module and a loss module; the gating mechanism module takes the original image as input, and the gating mechanism enables multiple differentiable image processing modules to perform relative weighting concurrently to enhance the image under adverse environmental conditions and output the image; the input part of the global module is composed of the feature image and the original image obtained above, and the output obtains the final enhanced output; the input of the loss module is the enhanced output and the original image; the loss module is connected after the gating mechanism module; the weights of the model are continuously optimized through training; finally, the enhanced output obtained is used as the input of the target detection model to perform subsequent target recognition tasks.

2. A method for target recognition in complex environments based on deep learning according to claim 1, characterized in that: The gating mechanism module is further improved by multiple GDIPs through a multi-stage guided process for gradual image enhancement; while ensuring speed, the gating mechanism module proposes a variant of GDIP in order to balance accuracy and speed, thus eliminating the need for GDIP-based image enhancement during inference, thereby achieving higher throughput and feasible real-world deployment; By combining the GDIP block with each layer of the visual encoder, multi-level progressive image enhancement is achieved.

3. The method for target recognition in complex environments based on deep learning according to claim 1, characterized in that: The input of the global module consists of a key and a value, the key refers to the original image, and the value refers to the feature image. The global branch consists of a first attention module and a second attention module, wherein the attention modules are linearly connected. The attention modules adopt the ACmis self-attention mechanism, which combines convolution and self-attention aggregation, wherein 1x1 convolution is shared between the two, aiming to reduce computational overhead and integrate lightweight aggregation operations.

4. The method for target recognition in complex environments based on deep learning according to claim 1, characterized in that: The loss module consists of three parts: smoothing loss, structure-aware smoothing loss, and consistency loss. First, the gradient strength T of the original image is calculated according to the gradient strength calculation formula. s and noise reduction feature T D .

5. The method for target recognition in complex environments based on deep learning according to claim 4, characterized in that: The formula for the smoothing loss is L smooth =||T D ||, where ‖·‖ refers to the L2 norm.

6. The method for target recognition in complex environments based on deep learning according to claim 4, characterized in that: The formula of the structure-aware smoothness loss is: Among them, λ is a hyperparameter, which refers to the structure perception strength, Refers to element-wise multiplication.

Images