Hotel entrance monitoring image enhancement method

Abstract

The invention provides a hotel entrance monitoring image enhancement method, and relates to the field of image enhancement. The invention provides a hotel entrance monitoring image multi-scale optimization processing flow. Comprising the steps of manufacturing a hotel entrance monitoring image data set, constructing a first-class high-dimensional feature processing module, constructing a second-class high-dimensional feature processing module, constructing a third-class high-dimensional feature processing module, constructing a hotel entrance monitoring image multi-scale optimization processing model, and training and using the hotel entrance monitoring image multi-scale optimization processing model. Meanwhile, a hotel entrance monitoring image multi-scale optimization processing model is provided, a first-class high-dimensional feature processing module uses Fourier transform to carry out frequency domain transformation processing on features, and a second-class high-dimensional feature processing module uses a random index array to carry out space pixel uniform arrangement. Then a corresponding inverse random index array is used for carrying out space pixel reduction, and the three types of high-dimensional feature processing modules carry out feature processing through convolution and channel attention.

Description

Technical Field

[0001] The invention belongs to the field of image enhancement, and in particular relates to a hotel entrance monitoring image enhancement method. Background Art

[0002] With the development of intelligent security technology, the importance of monitoring systems in hotel security management is increasing. However, due to insufficient light, weather conditions or equipment performance, the monitoring images at the hotel entrance often have low clarity and cannot accurately capture details. This not only affects the monitoring effect of managers, but may also pose a hidden danger to the safety of the hotel. Therefore, improving the clarity of monitoring images has become an important issue.

[0003] Image enhancement technology can effectively improve the visual quality of surveillance images, making originally blurry, low-resolution images clearer. This type of technology mainly includes brightness adjustment, contrast enhancement, noise suppression, and image sharpening. Through a series of algorithm optimizations, surveillance images can more realistically restore the on-site situation, help staff better identify suspicious persons or behaviors, and improve the overall security level of the hotel.

[0004] The use of deep learning technology for image enhancement has significant advantages. Unlike traditional image processing methods, deep learning can automatically learn complex features in images and perform precise optimization by building and training neural network models, so that monitoring images can still improve clarity under adverse conditions such as low light and blur. This method can automatically adapt to different scenarios, reduce the need for manual debugging, and show excellent results in noise suppression and detail recovery. The application of deep learning not only improves the visual quality of images, but also provides more accurate image information for intelligent monitoring, bringing efficient and reliable support for hotel security management. Summary of the invention

[0005] The invention provides a hotel entrance monitoring image enhancement method, aiming to propose a multi-scale optimization processing model for the hotel entrance monitoring image, wherein a first type of high-dimensional feature processing module uses Fourier transform and inverse Fourier transform to perform frequency domain transformation processing on features, and extracts high-level frequency domain information through deep convolution and low-rank self-attention, a second type of high-dimensional feature processing module uses a random index array to uniformly arrange spatial pixels, thereby uniformizing the dark area of ​​the image as a whole, and performs feature processing through deep convolution and spatial attention, and then uses the corresponding inverse random index array to restore the spatial pixels, a third type of high-dimensional feature processing module also uses a random index array and a corresponding inverse random index array to perform spatial pixel processing, and performs feature processing through convolution and channel attention, and by combining multi-scale feature processing and the first, second and third types of high-dimensional feature processing modules, an optimized hotel entrance monitoring image is obtained, and better hotel entrance monitoring image quality is achieved.

[0006] The present invention aims to propose a multi-scale optimization processing model for hotel entrance monitoring images and provide a hotel entrance monitoring image enhancement method, comprising the following steps: S1. Preparation of hotel entrance monitoring image dataset: using high-resolution image acquisition equipment to obtain multiple hotel entrance monitoring images with rich details as multiple optimized hotel entrance monitoring images, blurring the multiple optimized hotel entrance monitoring images, including applying Gaussian blur, mean blur, motion blur and downsampling to reduce the details and quality of the multiple optimized hotel entrance monitoring images, obtaining multiple standard hotel entrance monitoring images, and corresponding to the multiple optimized hotel entrance monitoring images, forming a hotel entrance monitoring image dataset; S2, construct a class of high-dimensional feature processing modules, including Fourier transform and inverse Fourier transform; S3, build two types of high-dimensional feature processing modules, including index array, spatial attention and deep convolution; S4. Construct three types of high-dimensional feature processing modules, including index array, channel attention and deep convolution; S5. Construct a multi-scale optimization processing model for hotel entrance monitoring images, including input, downsampling, feature generation convolution, layer normalization, one type of high-dimensional feature processing module, two types of high-dimensional feature processing module, three types of high-dimensional feature processing module, transposed convolution and image generation layer; S6. Training and use of the multi-scale optimization processing model for hotel entrance surveillance images. Use the hotel entrance surveillance image dataset to train the multi-scale optimization processing model for hotel entrance surveillance images. After the training is completed, input the standard hotel entrance surveillance image to be processed into the multi-scale optimization processing model for hotel entrance surveillance images to obtain the optimized hotel entrance surveillance image.

[0007] Preferably, in step S2, for a class of high-dimensional feature processing modules, given the input feature , , , and Represents the given input features The height, width and channel of , , , represents Fourier transform, represents depthwise convolution, represents low-rank self-attention, stands for inverse Fourier transform, which converts the features Transformed to high-dimensional frequency domain space, and optimized by deep convolution and low-rank attention, the features are Combine spatiotemporal information with frequency domain information to fully capture input features Global structure and local details.

[0008] Preferably, in step S2, for a class of high-dimensional feature processing modules, the input features are first converted to the frequency domain through Fourier transform, and then high-level features are extracted in the frequency domain using a deep convolutional network. Subsequently, a low-rank self-attention mechanism is introduced to capture the global dependencies between frequency domain features and enhance the representation capability of the features. Finally, the processed frequency domain features are converted back to the original space through an inverse Fourier transform.

[0009] Preferably, in step S3, for the second type of high-dimensional feature processing module, given the input feature , , , and Represents the given input features The height, width and channels of the input features are then given Input into the first MLP to get features , ,feature And given input features The dimensions are consistent. MLP is a feedforward neural network. The basic structure consists of an input layer, a hidden layer, and an output layer. Generate a random index array B from the non-channel dimension of , that is, the dimension of index array B is , each value in index array B is in Within the range of , and without repeated indexes, the feature The elements in are rearranged according to the new positions specified by the index array B to generate features , the dimension of each element is , , and create a new index array C with dimensions , used to record features Each element in the feature The original position in the Input into deep convolution to get features , , then the feature Input into spatial attention to get features , , then the feature Rearrange the features according to the new positions specified by the index array C , , then the feature Input into the second MLP to get features , ,feature And given input features The dimensions are consistent, the characteristics It is the output of the second-class high-dimensional feature processing module.

[0010] Preferably, in step S3, for the second type of high-dimensional feature processing module, a uniform arrangement of spatial pixels is achieved through a random index array, thereby improving the dark area, and deep convolution and spatial attention mechanisms are used for advanced feature extraction and processing, and finally the spatial pixels are accurately restored through the corresponding inverse random index array.

[0011] Preferably, in step S4, for the three types of high-dimensional feature processing modules, given the input feature , , , and Represents the given input features The height, width and channels of the input features are then given Input into the first MLP to get features , ,feature And given input features The dimensions are consistent. MLP is a feedforward neural network. The basic structure consists of an input layer, a hidden layer, and an output layer. Generate a random index array B from the non-channel dimension of , that is, the dimension of index array B is , each value in index array B is in Within the range of , and without repeated indexes, the feature The elements in are rearranged according to the new positions specified by the index array B to generate features , the dimension of each element is , , and create a new index array C with dimensions , used to record features Each element in the feature The original position in the Input into 3×3 convolution to get features , , then the feature Input into channel attention to get features , , then the feature Rearrange the features according to the new positions specified by the index array C , , then the feature Input into the second MLP to get features , ,feature And given input features The dimensions are consistent, the characteristics It is the output of three types of high-dimensional feature processing modules.

[0012] Preferably, in step S4, for the three types of high-dimensional feature processing modules, similar to the two types of high-dimensional feature processing modules, uniform arrangement of spatial pixels is achieved through random index arrays, thereby improving dark areas, and convolution and channel attention mechanisms are used for advanced feature extraction and processing, and finally the spatial pixels are accurately restored through the corresponding inverse random index arrays.

[0013] Preferably, in step S5, for the multi-scale optimization processing model of hotel entrance monitoring images, a standard hotel entrance monitoring image is input , , , and 3 represent standard hotel entrance surveillance images height, width and passageway, standard hotel entrance surveillance image Downsample by 2 times to get the image , , and yes and half of the image Downsample by 2 times to get the image , , and yes and A quarter of the standard hotel entrance monitoring image is generated using a 3×3 convolution as the feature generation convolution. Input to the first feature to generate convolution to get the feature , , , and Representative features The height, width and channels of the image Input to the second feature to generate convolution to get the feature , , , and Representative features The height, width and channels of the image Input to the third feature to generate convolution to get the feature , , , and Representative features The height, width and channel of Get Features , , , The representative layer is normalized. represents element-by-element addition, Represents a class of high-dimensional feature processing modules, Represents the second-class high-dimensional feature processing module, which transforms the features Input to the first transposed convolution to get features , , the feature and Features Add element by element to get features , , through the characteristics and Features Get Features , , , The representative layer is normalized. represents element-by-element addition, Represents a class of high-dimensional feature processing modules, Represents the second-class high-dimensional feature processing module, which transforms the features Input to the second transposed convolution to get features , , the feature and Features Add element by element to get features , , through the characteristics and Features Get Features , , , The representative layer is normalized. represents element-by-element addition, Represents a class of high-dimensional feature processing modules, Represents three types of high-dimensional feature processing modules, Input to the image generation layer to get the optimized hotel entrance monitoring image , , the image generation layer is implemented by 3×3 convolution, optimizing the hotel entrance monitoring image It is the output of the multi-scale optimization processing model of the hotel entrance surveillance image. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The present invention provides a multi-scale optimization processing flow chart of hotel entrance monitoring images.

[0015] Figure 2 It is a structural diagram of the second type of high-dimensional feature processing module provided by the present invention.

[0016] Figure 3 This is a standard hotel entrance monitoring image provided by the present invention.

[0017] Figure 4 This is the optimized hotel entrance monitoring image provided by the present invention. DETAILED DESCRIPTION

[0018] The present invention aims to propose a hotel entrance monitoring image enhancement method, and proposes a multi-scale optimization processing model for hotel entrance monitoring images, wherein a type of high-dimensional feature processing module uses Fourier transform and inverse Fourier transform to perform frequency domain transformation processing on features, and extracts advanced frequency domain information through deep convolution and low-rank self-attention, a type of high-dimensional feature processing module uses a random index array to uniformly arrange spatial pixels, thereby uniformly arranging the dark area of ​​the image as a whole, and performs feature processing through deep convolution and spatial attention, and then uses the corresponding inverse random index array to restore spatial pixels, a type of high-dimensional feature processing module also uses a random index array and a corresponding inverse random index array to perform spatial pixel processing, and performs feature processing through convolution and channel attention, and by combining multi-scale feature processing and the first, second, and third types of high-dimensional feature processing modules, an optimized hotel entrance monitoring image is obtained, and better hotel entrance monitoring image quality is achieved.

[0019] See also Figure 1 As shown, a hotel entrance monitoring image enhancement method in an embodiment of the present application: S1. The hotel entrance monitoring image dataset is prepared by using a high-resolution image acquisition device to obtain 450 detailed hotel entrance monitoring images as 450 optimized hotel entrance monitoring images. The height and width of each optimized hotel entrance monitoring image are 1280. The 450 optimized hotel entrance monitoring images are blurred, including applying Gaussian blur, mean blur, motion blur and downsampling to reduce the details and quality of the 450 optimized hotel entrance monitoring images, and 450 standard hotel entrance monitoring images are obtained. The height and width of each standard hotel entrance monitoring image are 1280, and the corresponding 450 optimized hotel entrance monitoring images form a hotel entrance monitoring image dataset, which contains a total of 450 training pairs; S2, construct a class of high-dimensional feature processing modules, including Fourier transform and inverse Fourier transform; S3, build two types of high-dimensional feature processing modules, including index array, spatial attention and deep convolution; S4. Construct three types of high-dimensional feature processing modules, including index array, channel attention and deep convolution; S5. Construct a multi-scale optimization processing model for hotel entrance monitoring images, including input, downsampling, feature generation convolution, layer normalization, one type of high-dimensional feature processing module, two types of high-dimensional feature processing module, three types of high-dimensional feature processing module, transposed convolution and image generation layer; S6. Training and use of the multi-scale optimization processing model for hotel entrance surveillance images. The hotel entrance surveillance image dataset is used to train the multi-scale optimization processing model for hotel entrance surveillance images. 350 training pairs are used for training. After the training is completed, a single standard hotel entrance surveillance image to be processed is input into the multi-scale optimization processing model for hotel entrance surveillance images among the remaining 100 training pairs to obtain a single optimized hotel entrance surveillance image.

[0020] Further, in step S2, for a Similar to high-dimensional feature processing module, given the input features , , , and Represents the given input features The height, width and channel of , , , represents Fourier transform, represents depthwise convolution, represents low-rank self-attention, stands for inverse Fourier transform, which converts the features Transformed to high-dimensional frequency domain space, and optimized by deep convolution and low-rank attention, the features are Combine spatiotemporal information with frequency domain information to fully capture input features Global structure and local details.

[0021] Further, in step S3, for the second type of high-dimensional feature processing module, its structure is as follows Figure 2 As shown, given the input features , , , and Represents the given input features The height, width and channels of the input features are then given Input into the first MLP to get features , ,feature And given input features The dimensions are consistent. MLP is a feedforward neural network. The basic structure consists of an input layer, a hidden layer, and an output layer. Generate a random index array B from the non-channel dimension of , that is, the dimension of index array B is , each value in index array B is in Within the range of , and without repeated indexes, the feature The elements in are rearranged according to the new positions specified by the index array B to generate features , the dimension of each element is , , and create a new index array C with dimensions , used to record features Each element in the feature The original position in the Input into deep convolution to get features , , then the feature Input into spatial attention to get features , , then the feature Rearrange the features according to the new positions specified by the index array C , , then the feature Input into the second MLP to get features , ,feature And given input features The dimensions are consistent, the characteristics It is the output of the second-class high-dimensional feature processing module.

[0022] Further, in step S4, for the three types of high-dimensional feature processing modules, given the input features , , , and Represents the given input features The height, width and channels of the input features are then given Input into the first MLP to get features , ,feature And given input features The dimensions are consistent. MLP is a feedforward neural network. The basic structure consists of an input layer, a hidden layer, and an output layer. Generate a random index array B from the non-channel dimension of , that is, the dimension of index array B is , each value in index array B is in Within the range of , and without repeated indexes, the feature The elements in are rearranged according to the new positions specified by the index array B to generate features , the dimension of each element is , , and create a new index array C with dimensions , used to record features Each element in the feature The original position in the Input into 3×3 convolution to get features , , then the feature Input into channel attention to get features , , then the feature Rearrange the features according to the new positions specified by the index array C , , then the feature Input into the second MLP to get features , ,feature And given input features The dimensions are consistent, the characteristics It is the output of three types of high-dimensional feature processing modules.

[0023] Further, in step S5, for the multi-scale optimization processing model of hotel entrance monitoring images, the standard hotel entrance monitoring image is input , , 1280, 1280 and 3 represent standard hotel entrance surveillance images height, width and passageway, standard hotel entrance surveillance image Downsample by 2 times to get the image , , the image Downsample by 2 times to get the image , , using 3×3 convolution as the feature generation convolution, the standard hotel entrance monitoring image Input to the first feature to generate convolution to get the feature , , 64, 64 and 32 represent the characteristics The height, width and channels of the image Input to the second feature to generate convolution to get the feature , , 32, 32 and 64 represent the characteristics The height, width and channels of the image Input to the third feature to generate convolution to get the feature , , 16, 16 and 128 represent features The height, width and channel of Get Features , , , The representative layer is normalized. represents element-by-element addition, Represents a class of high-dimensional feature processing modules, Represents the second-class high-dimensional feature processing module, which transforms the features Input to the first transposed convolution to get features , , the feature and Features Add element by element to get features , , through the characteristics and Features Get Features , , , The representative layer is normalized. represents element-by-element addition, Represents a class of high-dimensional feature processing modules, Represents the second-class high-dimensional feature processing module, which transforms the features Input to the second transposed convolution to get features , , the feature and Features Add element by element to get features , , through the characteristics and Features Get Features , , , The representative layer is normalized. represents element-by-element addition, Represents a class of high-dimensional feature processing modules, Represents three types of high-dimensional feature processing modules, Input to the image generation layer to get the optimized hotel entrance monitoring image , , the image generation layer is implemented by 3×3 convolution, optimizing the hotel entrance monitoring image It is the output of the multi-scale optimization processing model of the hotel entrance surveillance image. Compared with the standard hotel entrance surveillance image, the optimized hotel entrance surveillance image has better image quality and more details.

[0024] Furthermore, in step S5, the multi-scale optimization processing model of the hotel entrance surveillance image is implemented based on the PyTorch framework and trained using the NVIDIA GeForce RTX 4080 GPU with 16GB of video memory. The Adam optimizer is used, the learning rate is initialized to 0.0001, and an exponential annealing scheme is used, and the final learning rate is gradually reduced to 0.00001.

[0025] further, Figure 3 This is a standard hotel entrance surveillance image. Figure 4 This is an optimized hotel entrance surveillance image. Figure 3 and Figure 4 The resolution is 1280×1280. In step S6, after training is completed, input Figure 3 The standard hotel entrance monitoring image to the hotel entrance monitoring image multi-scale optimization processing model is shown, and the Figure 4 Optimized hotel entrance surveillance image shown.

[0026] The above are only preferred embodiments of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A hotel entrance monitoring image enhancement method, characterized in that: The following steps are involved: S1. Preparation of hotel entrance monitoring image dataset: using high-resolution image acquisition equipment to obtain multiple hotel entrance monitoring images with rich details as multiple optimized hotel entrance monitoring images, blurring the multiple optimized hotel entrance monitoring images, including applying Gaussian blur, mean blur, motion blur and downsampling to reduce the details and quality of the multiple optimized hotel entrance monitoring images, obtaining multiple standard hotel entrance monitoring images, and corresponding to the multiple optimized hotel entrance monitoring images, forming a hotel entrance monitoring image dataset; S2, construct a class of high-dimensional feature processing modules, including Fourier transform and inverse Fourier transform; S3, build two types of high-dimensional feature processing modules, including index array, spatial attention and deep convolution; S4. Construct three types of high-dimensional feature processing modules, including index array, channel attention and deep convolution; S5. Construct a multi-scale optimization processing model for hotel entrance monitoring images, including input, downsampling, feature generation convolution, layer normalization, one type of high-dimensional feature processing module, two types of high-dimensional feature processing module, three types of high-dimensional feature processing module, transposed convolution and image generation layer; S6. Training and use of the multi-scale optimization processing model for hotel entrance surveillance images. Use the hotel entrance surveillance image dataset to train the multi-scale optimization processing model for hotel entrance surveillance images. After the training is completed, input the standard hotel entrance surveillance image to be processed into the multi-scale optimization processing model for hotel entrance surveillance images to obtain the optimized hotel entrance surveillance image.

2. A hotel entrance monitoring image enhancement method according to claim 1, characterized in that: In step S2, for a class of high-dimensional feature processing modules, given the input feature , , , and Represents the given input features The height, width and channel of , and then get the features , , , represents Fourier transform, represents depthwise convolution, represents low-rank self-attention, stands for inverse Fourier transform, which converts the features Transformed to high-dimensional frequency domain space, and optimized by deep convolution and low-rank attention, the features are Combine spatiotemporal information with frequency domain information to fully capture input features Global structure and local details.

3. A hotel entrance monitoring image enhancement method according to claim 1, characterized in that: In step S3, for the second-class high-dimensional feature processing module, given the input feature , , , and Represents the given input features The height, width and channels of the input features are then given Input into the first MLP to get features , ,feature And given input features The dimensions are consistent. MLP is a feedforward neural network. The basic structure consists of an input layer, a hidden layer, and an output layer. Generate a random index array B from the non-channel dimension of , that is, the dimension of index array B is , each value in index array B is in Within the range of , and without repeated indexes, the feature The elements in are rearranged according to the new positions specified by the index array B to generate features , the dimension of each element is , , and create a new index array C with dimensions , used to record features Each element in the feature The original position in the Input into deep convolution to get features , , then the feature Input into spatial attention to get features , , then the feature Rearrange the features according to the new positions specified by the index array C , , then the feature Input into the second MLP to get features , ,feature And given input features The dimensions are consistent, the characteristics It is the output of the second-class high-dimensional feature processing module.

4. A hotel entrance monitoring image enhancement method according to claim 1, characterized in that: In step S4, for the three types of high-dimensional feature processing modules, given the input features , , , and Represents the given input features The height, width and channels of the input features are then given Input into the first MLP to get features , ,feature And given input features The dimensions are consistent. MLP is a feedforward neural network. The basic structure consists of an input layer, a hidden layer, and an output layer. Generate a random index array B from the non-channel dimension of , that is, the dimension of index array B is , each value in index array B is in Within the range of , and without repeated indexes, the feature The elements in are rearranged according to the new positions specified by the index array B to generate features , the dimension of each element is , , and create a new index array C with dimensions , used to record features Each element in the feature The original position in the Input into 3×3 convolution to get features , , then the feature Input into channel attention to get features , , then the feature Rearrange the features according to the new positions specified by the index array C , , then the feature Input into the second MLP to get features , ,feature And given input features The dimensions are consistent, the characteristics It is the output of three types of high-dimensional feature processing modules.

5. A hotel entrance monitoring image enhancement method according to claim 1, characterized in that: In step S5, for the multi-scale optimization processing model of hotel entrance monitoring images, the standard hotel entrance monitoring image is input , , , and 3 represent standard hotel entrance surveillance images height, width and passageway, standard hotel entrance surveillance image Downsample by 2 times to get the image , , and yes and half of the image Downsample by 2 times to get the image , , and yes and A quarter of the standard hotel entrance monitoring image is generated using a 3×3 convolution as the feature generation convolution. Input to the first feature to generate convolution to get the feature , , , and Representative features The height, width and channels of the image Input to the second feature to generate convolution to get the feature , , , and Representative features The height, width and channels of the image Input to the third feature to generate convolution to get the feature , , , and Representative features The height, width and channel of Get Features , , , The representative layer is normalized. represents element-by-element addition, Represents a class of high-dimensional feature processing modules, Represents the second-class high-dimensional feature processing module, which transforms the features Input to the first transposed convolution to get features , , the feature and Features Add element by element to get features , , through the characteristics and Features Get Features , , , The representative layer is normalized. represents element-by-element addition, Represents a class of high-dimensional feature processing modules, Represents the second-class high-dimensional feature processing module, which transforms the features Input to the second transposed convolution to get features , , the feature and Features Add element by element to get features , , through the characteristics and Features Get Features , , , The representative layer is normalized. represents element-by-element addition, Represents a class of high-dimensional feature processing modules, Represents three types of high-dimensional feature processing modules, Input to the image generation layer to get the optimized hotel entrance monitoring image , , the image generation layer is implemented by 3×3 convolution, optimizing the hotel entrance monitoring image It is the output of the multi-scale optimization processing model of the hotel entrance surveillance image.

Images