Pedestrian image enhancement method and device based on multi-kernel feature fusion convolutional neural network, and medium

Abstract

The invention provides a pedestrian image enhancement method and device based on a multi-kernel feature fusion convolutional neural network, and a medium, and belongs to the technical field of computer vision. The method comprises the steps that pedestrian images shot under the severe weather condition and the shielding condition are collected to serve as an original image data set, and the original image data set is marked to serve as a training data set; establishing a pedestrian image enhancement model based on a multi-kernel feature fusion convolutional neural network; training a pedestrian image enhancement model through a training set, inputting the training data set into the neural network model, and training the network by adopting a supervised learning mode until a training termination condition is met; and deploying the trained pedestrian image enhancement model to the front end of the target detection network, and enhancing the input image of the target detection network. According to the invention, the definition and recognition accuracy of pedestrian imaging are improved, and the generalization ability and robustness of the model are improved.

Description

Technical Field

[0001] The present invention relates to a pedestrian image enhancement method, device and medium based on a multi-core feature fusion convolutional neural network, and belongs to the technical field of computer vision. Background Art

[0002] With the rapid development of artificial intelligence and computer vision technology, pedestrian recognition technology has been widely used in intelligent monitoring, autonomous driving, smart cities and other fields. Pedestrian recognition aims to detect and identify specific individuals by analyzing the characteristics of pedestrians in images or videos. Traditional pedestrian recognition methods usually rely on manually designed features such as edges, textures, colors, etc. These methods perform well in simple scenes, but it is difficult to achieve satisfactory results in complex environments (such as lighting changes, occlusion, fog, rain).

[0003] In recent years, the rise of convolutional neural networks (CNNs) has brought revolutionary changes to pedestrian recognition. By automatically learning multi-level features in images, CNNs have made significant progress in pedestrian detection and recognition. Pedestrian recognition is often affected by different lighting conditions and complex background environments. Traditional CNNs have difficulty adapting to these changes, especially when lighting conditions change drastically, the stability and accuracy of the model will be significantly reduced. In actual application scenarios, pedestrians are often partially occluded (such as by vehicles, objects or other pedestrians). Traditional CNNs are usually unable to effectively handle such partially visible situations, especially unable to extract enough useful features from the non-occluded parts, resulting in reduced recognition accuracy. Although CNNs have shown high accuracy in pedestrian recognition, they consume a lot of computing and storage resources. Especially for embedded devices or real-time systems, the complexity of traditional CNNs may lead to reduced recognition speed or failure to meet real-time requirements.

[0004] The Chinese invention patent with application number CN202410828405.8 discloses a pedestrian recognition method and device based on multi-core fusion of global features and head-shoulder features, which relates to the field of image recognition, including: using a trained pedestrian head-shoulder detection model to perform head-shoulder detection on a pedestrian image to obtain a pedestrian head-shoulder image; in the pedestrian recognition model, the pedestrian image and the pedestrian head-shoulder image are respectively input into the global feature extraction branch and the head-shoulder feature extraction branch to obtain a global feature vector and a head-shoulder feature vector and input them into a multi-core fusion module for fusion to obtain a fused feature vector, and a loss function is constructed according to the global feature vector, the head-shoulder feature vector and the fused feature vector to train the pedestrian recognition model to obtain a trained pedestrian recognition model; the pedestrian image to be identified and its corresponding pedestrian head-shoulder image are input into the trained pedestrian recognition model to obtain the corresponding fused feature vector, and then pedestrian recognition is performed. The present invention solves the problem of large image feature differences and low accuracy under fisheye cameras. However, the pedestrian recognition accuracy of the model for complex scenes is not high.

[0005] The application number is CN202410874053.X. The present invention proposes a visual language pedestrian re-identification network method and system based on dynamic attention. The method includes obtaining foreground features in the original image, segmenting it, and using the segmented foreground features to perform foreground perception processing on the original image to obtain a foreground perception image, sending the original image and the foreground perception image into the image encoder to obtain global image features, inputting the question text and the original image into the text encoder to generate text features, calculating the attention weights of the global image features and text features, and dynamically allocating weights according to the importance between the global image features and the text features to obtain the attention weighted features of the two modalities, and splicing them to obtain integrated features. The present invention dynamically allocates weights according to the importance between the global image features and the text features, overcomes the problem of weight allocation bias in the interaction between image and text information, and enables the model to correctly focus on useful information. The patent uses a complex model to handle pedestrian recognition tasks, resulting in large model parameters.

[0006] In the process of implementing the embodiments of the present disclosure, it is found that there are at least the following problems in the related art: Occlusion handling: Pedestrian recognition systems need to be able to accurately identify pedestrians even when they are partially occluded. For example, a pedestrian may be partially occluded by other objects (such as cars or other pedestrians), which makes it impossible to obtain complete features. Solving this problem usually requires using deep learning methods to learn features in partial occlusion situations, or combining data from multiple sensors.

[0007] Lighting and weather conditions: Different lighting and weather conditions (such as strong light, shadows, rain, snow, etc.) can affect the performance of the recognition system. Under these conditions, the quality and visibility of the image may be affected, resulting in reduced recognition accuracy. To overcome these challenges, the system usually needs to perform image enhancement processing or use adaptive algorithms to handle different environmental conditions.

[0008] Model complexity: Deep learning models usually need to process high-dimensional features to capture the complex characteristics of pedestrians. However, the use of high-dimensional features often leads to too many model parameters, which increases computational overhead and memory requirements. This makes model training and reasoning more difficult, especially in resource-constrained environments. Summary of the invention

[0009] The purpose of the present invention is to provide a pedestrian image enhancement method, device and medium based on a multi-core feature fusion convolutional neural network, which enhances the perception of pedestrian edges and contrast by extracting features at different levels and performing feature fusion, and improves the generalization ability and robustness of the model.

[0010] In order to achieve the above object, the present invention is implemented through the following technical solutions: A pedestrian image enhancement method based on a multi-core feature fusion convolutional neural network includes the following steps: Pedestrian images taken under adverse weather conditions and occlusion conditions are collected as original image datasets, and the original image datasets are labeled as training datasets; A pedestrian image enhancement model based on a multi-core feature fusion convolutional neural network is established, wherein the pedestrian image enhancement model sequentially includes a multi-core pooling layer, a channel convolution layer, a channel fusion layer, a multi-core convolution layer and a fully connected layer; Training a pedestrian image enhancement model using a training set, inputting the training data set into the neural network model, and training the network using a supervised learning method until a training termination condition is met; The trained pedestrian image enhancement model is deployed to the front end of the target detection network to enhance the input image of the target detection network.

[0011] Preferably, the original image dataset includes a RainyCityscapes dataset and a Foggy Zurich dataset.

[0012] Preferably, the window scales of the multi-core pooling layer and the multi-core convolutional layer of the pedestrian image enhancement model based on the multi-core feature fusion convolutional neural network include 3x3, 5x5 and 7x7; the multi-core pooling layer includes average pooling and maximum pooling.

[0013] Preferably, the pedestrian image enhancement model processes the image as follows: The original image dataset is input into the multi-core pooling layer, and the pooling operation is performed through the 3x3 pooling layer, the 5x5 pooling layer, and the 7x7 pooling layer to capture multi-scale features. The output features of each pooling layer are respectively extracted through the channel convolution layer to extract the feature information of different channels; The channel fusion layer adaptively assigns weights to each channel and fuses the features of each channel; The fused features are input into the multi-core convolution layer, and convolution operations are performed through the 3x3 convolution layer, 5x5 convolution layer, and 7x7 convolution layer respectively to capture the features of pedestrians in the feature image; The features output by each convolutional layer are output through the fully connected layer to enhance the pedestrian image.

[0014] Preferably, each channel in the channel convolution layer has an independent convolution kernel, and the convolution kernels perform convolution operations with the corresponding channels of the input image respectively; the calculation process is as follows: , in, represents the output of the channel convolution layer, whose number of channels is consistent with the input of this layer. represents the convolution kernel, ReLU activation function is shown. Represents the output of the multi-core pooling layer.

[0015] Preferably, the channel fusion layer processes the input features as follows: For input features Each channel performs independent full-connection calculations, and the calculation process is as follows: , in, Indicates the calculated The weight score of each channel, and Represents the corresponding weight matrix and bias vector; The calculated scores are multiplied by their corresponding channels and added together. The specific formula is as follows: , in, Represents the output of the channel fusion layer.

[0016] Preferably, the multi-core convolution layer processes the input features as follows: right Use the size of The convolution kernel is calculated, and the specific formula is as follows: , in, Indicates that the convolution kernel size is The output when Indicates size The convolution kernel, Represents the Sigmoid activation function; right Perform feature fusion: Flatten into one-dimensional vectors , and then forward calculation is performed through the fully connected layer. The formula is as follows: , in, Represents the weight of each channel, , Represents the learnable weight matrix; Multiply each channel by the corresponding weight and add them together to get the final output. The formula is as follows: , in, Represents the enhanced image obtained by the final output.

[0017] Preferably, the pedestrian image enhancement model training process adopts a mean square error loss function as a training target, and performs gradient update through an Adam optimization algorithm.

[0018] A pedestrian image enhancement device based on a multi-core feature fusion convolutional neural network comprises a processor and a memory storing program instructions, wherein the processor is configured to execute the pedestrian image enhancement method based on the multi-core feature fusion convolutional neural network when running the program instructions.

[0019] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the pedestrian image enhancement method based on a multi-core feature fusion convolutional neural network.

[0020] The advantages of the present invention are: Strong multi-core feature fusion capability Compared with single-core convolutional neural networks, multi-core CNNs can process features of different scales at the same time, from local to global. By extracting features at different levels and fusing features, the ability to perceive the edges and contrast of pedestrians is enhanced. This multi-scale processing mechanism enables the network to better adapt to pedestrians of different scales and shapes, improving the clarity and recognition accuracy of pedestrian imaging. Multi-core CNNs can also enhance the robustness of the network, reduce the risk of overfitting, and improve the adaptability of the model under different image transformations.

[0021] Strong feature expression ability The channel convolution layer can apply convolution operations independently on each channel, thereby better capturing the feature information of different channels in the image and enhancing the perception of pedestrians. The channel fusion layer combines the features of different channels to produce a richer and more comprehensive feature representation, which can more accurately describe the morphology of pedestrians. By reducing the redundancy between features, the feature representation is more representative and stable, improving the generalization and robustness of the model.

[0022] Adapting to complex environments This method can effectively handle pedestrian recognition problems under partial occlusion and complex lighting conditions, and can also achieve high enhancement effects on blurry or low-quality images taken in harsh environments. Through comprehensive data collection and annotation, the method covers a variety of complex environments and conditions and can meet the needs of actual application scenarios. Compared with traditional methods, this method performs better in complex environments and has strong practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0024] Figure 1 It is a schematic diagram of the process of the present invention.

[0025] Figure 2 The overall block diagram of the pedestrian image enhancement model.

[0026] Figure 3 This is the structure diagram of the pedestrian image enhancement model. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] Example 1 like Figure 1 As shown, a pedestrian image enhancement method based on a multi-core feature fusion convolutional neural network includes the following steps: S1: Pedestrian images taken under adverse weather conditions and occlusion conditions are collected as the original image dataset, and the original image dataset is labeled as the training dataset.

[0029] S2: Establish a pedestrian image enhancement model based on a multi-core feature fusion convolutional neural network, wherein the pedestrian image enhancement model includes a multi-core pooling layer, a channel convolution layer, a channel fusion layer, a multi-core convolution layer and a fully connected layer in sequence.

[0030] S3: Train the pedestrian image enhancement model through the training set, input the training data set into the neural network model, and train the network by supervised learning until the training termination condition is met.

[0031] S4: Deploy the trained pedestrian image enhancement model to the front end of the target detection network to enhance the input image of the target detection network.

[0032] As a refinement of the above embodiment, in order to train and verify the pedestrian recognition image enhancement method based on multi-core feature fusion convolutional neural network, we conducted comprehensive data collection and labeling work on blurry original pictures taken under severe weather conditions. The data collection and processing ensures that various complex environments and conditions are covered, and the enhancement effect is verified through high-quality output. The image data sources cover a variety of blurry image acquisition scenarios, focusing on low-quality images obtained under severe weather conditions and occlusion conditions.

[0033] The dataset collection methods include public datasets and self-collected datasets, as follows: Public datasets: We used public datasets that include pedestrian images taken in blurry or low-visibility environments, such as RainyCityscapes, Foggy Zurich, etc. These datasets were collected under adverse conditions such as rainy days, foggy days, and nights, and contain typical blur and occlusion problems.

[0034] Self-owned data collection: Pedestrian images collected by monitoring equipment and vehicle cameras under adverse weather conditions (such as heavy rain, dense fog, low light at night, etc.) and occlusion conditions are used to simulate the complexity and challenges in real application scenarios. In this way, a large number of original images are obtained. Then, a professional team uses advanced image enhancement technology to accurately repair the pedestrians in the blurred image, correcting the distorted edges, details and blurred areas to high-definition state. This process ensures that the pedestrians' outlines are clearly visible and the key features are preserved. The enhanced image is called the enhanced image, i.e. . Each original image and its corresponding enhanced image are taken as a pair of training samples, and they are paired to form a training data set. During the annotation process, the quality of the generated enhanced images is evaluated to ensure that they meet the expected standards in terms of contrast, clarity, etc. At the same time, multiple reviews are carried out to ensure the accuracy and consistency of the annotations. Through such data collection and annotation processes, a high-quality and diverse training data set is obtained, which is used to train the neural network model and realize the automatic enhancement of pedestrian images. Finally, the data set is obtained ,in, represents the original image, Indicates a high-quality image that has been enhanced by a professional.

[0035] As a refinement of the above embodiment, compared with the traditional single-core convolutional neural network, the use of a kernel convolutional neural network has significant advantages in pedestrian imaging. Traditional single-scale CNNs usually only focus on features of a specific scale when processing images, and it is difficult to fully capture information of different scales in the image. Multi-core CNNs can process features of different scales at the same time, from local to global, and enhance the perception of the edges and contrast of pedestrians by extracting features at different levels and performing feature fusion, so that the network can more comprehensively understand and capture pedestrian information in the image. Such a multi-scale processing mechanism enables the network to better adapt to different environments and improves the clarity and recognition accuracy of pedestrian imaging. In addition, multi-core CNNs can also enhance the robustness of the network, reduce the risk of overfitting, and improve the adaptability of the model under different image transformations. In summary, multi-core convolutional neural networks have obvious advantages in pedestrian imaging, which can improve image quality, recognition accuracy and model robustness.

[0036] Optional, such as Figure 3 As shown, the pedestrian image enhancement model processes the image as follows: (1) The original image dataset is input into the multi-core pooling layer, and pooling operations are performed through 3x3 pooling layer, 5x5 pooling layer and 7x7 pooling layer to capture multi-scale features. The output features of each pooling layer are respectively passed through the channel convolution layer to extract the feature information of different channels.

[0037] Compared with the single-core pooling layer, the multi-core pooling layer can better improve the performance of the neural network. The single-core pooling layer usually uses a fixed pooling window size and stride to reduce the image size, which will lead to information loss and image blur. Especially in image enhancement, the single-core pooling layer cannot fully capture the information of different scales in the image, resulting in the loss of local details or global structural information of the pedestrian image. This will affect the clarity and recognition accuracy of the image enhancement. Therefore, the introduction of the multi-core pooling layer can better solve these problems and improve the quality and accuracy of pedestrian image enhancement.

[0038] The multi-core pooling layer is used for the input image The processing process is expressed as: , , in, and They represent the pooling kernels respectively. The output after average pooling and maximum pooling is The default value is {3, 5, 7}. By using border padding, the output length is guaranteed to be consistent after processing with different pooling kernels. Finally, the output of the multi-scale pooling layer is obtained. .

[0039] (2) The channel fusion layer adaptively assigns weights to each channel and fuses the features of each channel.

[0040] In pedestrian imaging enhancement, the channel convolution layer applies convolution operations independently on each channel, so that the network can better capture the feature information of different channels in the image, thereby enhancing the perception of pedestrians. At the same time, reducing the number of parameters improves the efficiency of the model, allowing the network to extract features and process images faster, accelerating the process of pedestrian image enhancement. In addition, the channel-by-channel convolution layer can also realize feature fusion and combination, producing richer and more diverse feature representations, thereby improving the clarity of imaging and recognition accuracy.

[0041] In channel convolution, each channel has an independent convolution kernel, which performs convolution operations with the corresponding channels of the input image. This design enables the network to learn different feature representations on each channel, thereby better capturing the diversity and complexity of pedestrian images.

[0042] The convolution kernel performs convolution operation with the corresponding channels of the input image respectively; the calculation process is as follows: , in, represents the output of the channel convolution layer, whose number of channels is consistent with the input of this layer. represents the convolution kernel, ReLU activation function is shown. Represents the output of the multi-core pooling layer.

[0043] (3) The fused features are input into the multi-core convolution layer, and convolution operations are performed through a 3x3 convolution layer, a 5x5 convolution layer, and a 7x7 convolution layer respectively to capture the features of pedestrians in the feature image.

[0044] Channel fusion after channel-by-channel convolution helps enrich feature representation and improve feature expression capabilities. Channel-by-channel convolution can extract the features of each channel, and channel fusion combines these features to produce a more diverse and comprehensive feature representation, thereby better capturing the information in the image. In addition, channel fusion can also reduce redundant information. By merging features from different channels, it reduces the redundancy between features and makes the features more representative and stable. Ultimately, the feature representation after channel fusion has better generalization ability, stronger adaptability, can be applied to different image enhancement tasks, and has better robustness.

[0045] In order to calculate the weight of each channel, this solution constructs a channel fusion layer to adaptively assign weights to each channel to ensure the integrity of the transmitted information. , an independent full connection calculation is performed on each channel, and the calculation process is as follows: , in, Indicates the calculated The weight score of each channel, and Denote the corresponding weight matrix and bias vector.

[0046] The calculated scores are multiplied by their corresponding channels and added together. The specific formula is as follows: , in, Represents the output of the channel fusion layer.

[0047] (4) The features output by each convolutional layer are output through the fully connected layer to produce an enhanced pedestrian image.

[0048] The benefit of using multi-core convolution for image enhancement is that it can overcome the limitations of single-scale convolution and more comprehensively capture the various features of pedestrians in the image. Using multi-scale convolution can effectively extract features for pedestrians in different environments, thereby enhancing the perception of pedestrians. By performing convolution operations at different scales, the network can simultaneously take into account the local details and global structure of the image to produce a richer and more comprehensive feature representation. Such a feature representation can more accurately describe the morphology of pedestrians, improve the clarity of imaging and recognition accuracy. In addition, multi-core convolution can also enhance the robustness of the network, making it better adapted to the pedestrian perception of the environment, thereby improving the effect and stability of pedestrian image enhancement.

[0049] Optionally, the multi-core convolution layer processes the input features as follows: right Use the size of The convolution kernel is calculated, and the specific formula is as follows: , in, Indicates that the convolution kernel size is The output when Indicates size The convolution kernel, Represents the Sigmoid activation function; right ( , , ) for feature fusion: Flatten into one-dimensional vectors ( , , ), and then forward calculation is performed through the fully connected layer. The formula is as follows: , , , in, , , express , , The three channels get weights, , , Represents the learnable weight matrix; Multiply each channel by the corresponding weight and add them together to get the final output. The formula is as follows: , in, Represents the enhanced image obtained by the final output.

[0050] As a refinement of the above embodiment, the model is based on a multi-scale convolutional neural network to build a novel pedestrian image enhancement model. By inputting the labeled training set into the model, the network is trained using a supervised learning method. Figure 2 As shown, the specific training process is as follows: S301: Data preprocessing; this model adopts an end-to-end structure. The input of the model is the original image. You only need to set the batch size n for each training.

[0051] S302: Define the loss function; use the mean square error loss function: , in, represents the number of samples, is the corresponding label, It is the enhanced pedestrian image.

[0052] S303: Gradient update; Adam is an adaptive learning rate optimization algorithm, commonly used for training neural networks. It combines the ideas of momentum method and adaptive learning rate adjustment, and dynamically adjusts the learning rate of each parameter during the gradient descent process. The main idea of ​​the Adam algorithm is to adaptively adjust the learning rate based on the first-order moment estimate (i.e., the average value of the gradient) and the second-order moment estimate (i.e., the average value of the square of the gradient) of each parameter. This can make the update direction of the parameters more accurate and stable, and accelerate the convergence process of the model. This model is trained using Adam.

[0053] S304: Training termination; One of the methods for terminating training is to set the number of iterations. Before training begins, a fixed number of iterations can be set in advance as the termination condition for training. When the model reaches the set number of iterations, training stops. This method is simple and intuitive, easy to implement, and can ensure that the model will not continue to iterate indefinitely during training, thereby avoiding the problem of overfitting. Set the number of iterations, and when the model reaches the set number of iterations, training stops.

[0054] S305: Model saving: At the end of training, the final model parameters are saved as a deployment model for pedestrian image enhancement.

[0055] As a refinement of the above embodiment, the trained enhancement model is deployed to the front end of various target detection networks (such as YOLO), so that the input of the detection network will be an enhanced high-quality image, which can improve the detection accuracy.

[0056] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A pedestrian image enhancement method based on multi-core feature fusion convolutional neural network, characterized in that: The method comprises the following steps: Pedestrian images taken under adverse weather conditions and occlusion conditions are collected as original image datasets, and the original image datasets are labeled as training datasets; A pedestrian image enhancement model based on a multi-core feature fusion convolutional neural network is established, wherein the pedestrian image enhancement model sequentially includes a multi-core pooling layer, a channel convolution layer, a channel fusion layer, a multi-core convolution layer and a fully connected layer; Training a pedestrian image enhancement model using a training set, inputting the training data set into the neural network model, and training the network using a supervised learning method until a training termination condition is met; The trained pedestrian image enhancement model is deployed to the front end of the target detection network to enhance the input image of the target detection network.

2. The pedestrian image enhancement method based on multi-core feature fusion convolutional neural network according to claim 1 is characterized in that: The original image datasets include the RainyCityscapes dataset and the Foggy Zurich dataset.

3. The pedestrian image enhancement method based on multi-core feature fusion convolutional neural network according to claim 1 is characterized in that: The window scales of the multi-core pooling layer and the multi-core convolutional layer of the pedestrian image enhancement model based on the multi-core feature fusion convolutional neural network include 3x3, 5x5 and 7x7; the multi-core pooling layer includes average pooling and maximum pooling.

4. The pedestrian image enhancement method based on multi-core feature fusion convolutional neural network according to claim 3 is characterized in that: The pedestrian image enhancement model processes the image as follows: The original image dataset is input into the multi-core pooling layer, and the pooling operation is performed through the 3x3 pooling layer, the 5x5 pooling layer, and the 7x7 pooling layer to capture multi-scale features. The output features of each pooling layer are respectively extracted through the channel convolution layer to extract the feature information of different channels; The channel fusion layer adaptively assigns weights to each channel and fuses the features of each channel; The fused features are input into the multi-core convolution layer, and convolution operations are performed through the 3x3 convolution layer, 5x5 convolution layer, and 7x7 convolution layer respectively to capture the features of pedestrians in the feature image; The features output by each convolutional layer are output through the fully connected layer to enhance the pedestrian image.

5. The pedestrian image enhancement method based on multi-core feature fusion convolutional neural network according to claim 4 is characterized in that: Each channel in the channel convolution layer has an independent convolution kernel, and the convolution kernel performs convolution operation with the corresponding channel of the input image respectively; the calculation process is as follows: ， in, represents the output of the channel convolution layer, whose number of channels is consistent with the input of this layer. represents the convolution kernel, ReLU activation function is shown. Represents the output of the multi-core pooling layer.

6. The pedestrian image enhancement method based on multi-core feature fusion convolutional neural network according to claim 5 is characterized in that: The channel fusion layer processes the input features as follows: For input features Each channel performs independent full-connection calculations, and the calculation process is as follows: ， in, Indicates the calculated The weight score of each channel, and represents the corresponding weight matrix and bias vector; The calculated scores are multiplied by their corresponding channels and added together. The specific formula is as follows: ， in, Represents the output of the channel fusion layer.

7. The pedestrian image enhancement method based on multi-core feature fusion convolutional neural network according to claim 5 is characterized in that: The multi-core convolution layer processes the input features as follows: right Use the size of The convolution kernel is calculated, and the specific formula is as follows: ， in, Indicates that the convolution kernel size is The output when Indicates size The convolution kernel, Represents the Sigmoid activation function; right Perform feature fusion: Flatten into one-dimensional vectors , and then forward calculation is performed through the fully connected layer. The formula is as follows: ， in, Represents the weight of each channel, , Represents the learnable weight matrix; Multiply each channel by the corresponding weight and add them together to get the final output. The formula is as follows: ， in, Represents the enhanced image obtained by the final output.

8. The pedestrian image enhancement method based on multi-core feature fusion convolutional neural network according to claim 5 is characterized in that: The pedestrian image enhancement model training process adopts the mean square error loss function as the training target and performs gradient update through the Adam optimization algorithm.

9. A pedestrian image enhancement device based on a multi-core feature fusion convolutional neural network, comprising a processor and a memory storing program instructions, characterized in that: The processor is configured to execute the pedestrian image enhancement method based on a multi-core feature fusion convolutional neural network as described in any one of claims 1-8 when running the program instructions.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, a pedestrian image enhancement method based on a multi-core feature fusion convolutional neural network as described in any one of claims 1 to 8 is implemented.

Images