Ultra-low operation complexity image enhancement neural network

Abstract

The invention is suitable for the technical field of image enhancement, and particularly relates to an ultra-low operation complexity image enhancement neural network, which comprises a coarse processing network, a crack removal network and a training layer, the coarse processing network is used for performing preliminary image enhancement on an input image, segmenting the input image into a plurality of sub-graphs, constructing a 3 * 3 convolution and a ReLU activation function, transferring the sub-graphs into the 3 * 3 convolution, extracting features in the sub-graphs, generating feature graphs in one-to-one correspondence with the sub-graphs, and embedding the ReLU activation function into the feature graphs; and creating a first deep convolution processing chain. By constructing the training layer, the coarse processing network and the crack removal network can be trained, the operation efficiency of the coarse processing network and the crack removal network is greatly improved, and by integrating the coarse processing network, the crack removal network and the training layer, the performance of the graph is enhanced while the operation complexity of the network is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of image enhancement, and in particular to an ultra-low computational complexity image enhancement neural network. Background Art

[0002] Image enhancement is one of the commonly used technologies in digital image processing. In recent years, compared with the traditional image enhancement method based on image grayscale processing, the image enhancement method based on neural network has achieved better technical results. However, due to the particularly large computational complexity of the network, and with the sufficient computing resources of GPU and hardware configuration on the computer side, the neural network focuses on improving image processing performance at the expense of increasing the complexity of the network, which makes it difficult to promote mobile applications.

[0003] Existing neural network-based image enhancement methods, such as AlexNet, VGG16, InceptionV3, and ResNet50 networks, have high image processing performance, but their computational complexity is extremely high, which limits their mobility on mobile terminals. Therefore, "how to significantly reduce the computational complexity of the network while maintaining image enhancement performance" is the technical problem that the present invention needs to solve. Summary of the invention

[0004] The purpose of the present invention is to provide an ultra-low computational complexity image enhancement neural network to solve the problem raised in the above background technology of "how to significantly reduce the computational complexity of the network while maintaining the image enhancement performance".

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An ultra-low computational complexity image enhancement neural network, the neural network comprising: a rough processing network, a crack removal network and a training layer;

[0007] The coarse processing network is used to perform preliminary image enhancement on the input image, divide the input image into several sub-images, construct a 3x3 convolution and ReLU activation function, transfer the sub-images to the 3x3 convolution, extract the features of the sub-images, and generate Figure 1 a corresponding feature map, and embed the ReLU activation function into the feature map, create a first deep convolution processing chain, wherein the first deep convolution processing chain consists of 15 depthwise separable convolutions, transfer the sub-image into the deep convolution processing chain, extract the details of the sub-image, and superimpose the details into the corresponding feature map, and splice all the sub-images to obtain a coarse processed image;

[0008] The crack removal network is used to eliminate the splicing cracks in the rough processed image, use 3x3 convolution and ReLU activation function to extract features of the rough processed image again, create a second deep convolution processing chain, wherein the second deep convolution processing chain consists of 5 depthwise separable convolutions, refine the extracted features, and obtain an output image after integration;

[0009] The training layer is used to train the rough processing network and the crack removal network, and to create training sets for the rough processing network and the crack removal network respectively, and to perform training.

[0010] Furthermore, the rough processing network includes:

[0011] A segmentation module is used to segment the input image into several sub-images;

[0012] The feature extraction module is used to construct a 3x3 convolution and ReLU activation function, transfer the sub-graph into the 3x3 convolution, extract the features of the sub-graph, and generate Figure 1 A corresponding feature map, and embed the ReLU activation function into the feature map.

[0013] Furthermore, the rough processing network also includes:

[0014] A feature refinement module, configured to create a first deep convolution processing chain, wherein the first deep convolution processing chain is composed of 15 depthwise separable convolutions, transfer the sub-image into the deep convolution processing chain, extract the details of the sub-image, and superimpose the details into the corresponding feature map;

[0015] The stitching module is used to stitch all the sub-images to obtain a roughly processed image.

[0016] Further, the crack removal network comprises:

[0017] A re-extraction module, used for re-extracting features from the roughly processed image using a 3x3 convolution and a ReLU activation function;

[0018] An elimination module, configured to create a second deep convolution processing chain, wherein the second deep convolution processing chain consists of five depthwise separable convolutions to refine the extracted features;

[0019] The output module is used to obtain the output image after integration.

[0020] Furthermore, the training layer includes:

[0021] A creation module, used to create training sets for the rough processing network and the crack removal network respectively;

[0022] Training module, used for training.

[0023] Furthermore, the creation module includes:

[0024] A definition unit is used to use the 91Images dataset in the existing data as original data, perform grayscale conversion on the original data to obtain a grayscale image, and extract 25 image blocks from each of the grayscale images with a step length of 20, which are defined as clean image blocks;

[0025] An adding unit, configured to add salt and pepper noise of random density to the clean image block to obtain a noisy image block;

[0026] The generation unit is used to establish a mapping between the noisy image block and the clean image block, generate a mapping pair, summarize all the mapping pairs, obtain a training set corresponding to the rough processing network, and define it as DtasetA.

[0027] Furthermore, the creation module also includes:

[0028] A selection unit is used to select Pascal VOC from an existing data set, wherein the total number of images of Pascal VOC is 17125, convert the Pascal VOC into a grayscale image, and number the image;

[0029] The insertion unit is used to insert random density salt and pepper noise into the numbered Pascal VOC grayscale image to obtain the training set corresponding to the crack removal network, which is defined as DatasetB.

[0030] Furthermore, the training layer also includes:

[0031] The evaluation module is used to evaluate the image enhancement effect of the rough processing network and the crack removal network.

[0032] Furthermore, the evaluation module includes:

[0033] A construction unit is used to select the BSD300 image set and add salt and pepper noise of random density to construct an evaluation image set;

[0034] The output unit is used to determine the processing order of the evaluation image set, and input the evaluation image set into the rough processing network and the crack removal network in sequence according to the processing order to output the enhanced image.

[0035] Furthermore, the evaluation module also includes:

[0036] The result unit is used to evaluate the signal-to-noise ratio using PSNR, evaluate the information loss using MSE, evaluate the computational complexity using MFLOPS, generate an evaluation table, and use the evaluation table to correct the rough processing network and the crack removal network.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] By constructing a coarse processing network, the sub-image can be segmented and combined with depthwise separable convolution, which greatly reduces the computational complexity of the image. By constructing a crack removal network, the splicing cracks can be repaired to ensure the integrity of the image. By constructing a training layer, the coarse processing network and the crack removal network can be trained, which greatly improves their operating efficiency. By integrating the coarse processing network, the crack removal network and the training layer, the performance of the graphics can be enhanced while reducing the computational complexity of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A processing flow chart of an ultra-low computational complexity image enhancement neural network provided by an embodiment of the present invention;

[0040] Figure 2 A structural diagram of the ultra-low computational complexity image enhancement neural network provided by an embodiment of the present invention;

[0041] Figure 3 A block diagram of the composition of a coarse processing network in an ultra-low computational complexity image enhancement neural network provided by an embodiment of the present invention;

[0042] Figure 4 A block diagram of the composition of a crack removal network in an ultra-low computational complexity image enhancement neural network provided by an embodiment of the present invention;

[0043] Figure 5 A block diagram of the composition of the training layer in the ultra-low computational complexity image enhancement neural network provided by an embodiment of the present invention;

[0044] Figure 6 A block diagram of the composition of a creation module in an ultra-low computational complexity image enhancement neural network provided by an embodiment of the present invention;

[0045] Figure 7 A flowchart of the DtasetA construction process in the ultra-low computational complexity image enhancement neural network provided by an embodiment of the present invention;

[0046] Figure 8 A block diagram of the composition of the evaluation module in the ultra-low computational complexity image enhancement neural network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0048] It is understood that the terms "first", "second", etc. used in this application may be used herein to describe various elements, but unless otherwise specified, these elements are not limited by these terms. These terms are only used to distinguish a first element from another element. For example, without departing from the scope of this application, a first ultra-low computational complexity image enhancement neural network script may be referred to as a second ultra-low computational complexity image enhancement neural network script, and similarly, a second ultra-low computational complexity image enhancement neural network script may be referred to as a first ultra-low computational complexity image enhancement neural network script.

[0049] In Example 1, Figure 1 and Figure 2 The composition of the ultra-low computational complexity image enhancement neural network provided by an embodiment of the present invention is shown and described in detail below:

[0050] An ultra-low computational complexity image enhancement neural network 1, comprising: a rough processing network, a crack removal network and a training layer;

[0051] The coarse processing network 11 is used to perform preliminary image enhancement on the input image, divide the input image into several sub-images, construct a 3x3 convolution and ReLU activation function, transfer the sub-images to the 3x3 convolution, extract the features of the sub-images, and generate Figure 1 A corresponding feature map is obtained, and the ReLU activation function is embedded into the feature map, and a first deep convolution processing chain is created, wherein the first deep convolution processing chain is composed of 15 depthwise separable convolutions, and the sub-image is transferred to the deep convolution processing chain, and the details of the sub-image are extracted, and the details are superimposed on the corresponding feature map, and all the sub-images are spliced ​​to obtain a coarsely processed image.

[0052] First, the input image is divided into several sub-images, of which 2×2 is common, and other sub-images such as 2×4, 4×2 and 4×4 can also be used; the smaller the sub-images are, the lower the computational complexity of the input image, but the more severe the stitching cracks are; then, the input image is subjected to preliminary feature extraction through the convolution layer, and the ReLU activation function is used to introduce nonlinearity, allowing the network to learn complex patterns and generate feature maps; the first deep convolution processing chain is created, where the first deep convolution processing chain is mainly used to extract more detailed features, and then the corresponding sub-images are obtained through 3x3 convolution; finally, all the sub-images are spliced ​​into an image with the same resolution as the input image.

[0053] The crack removal network 12 is used to eliminate the splicing cracks in the rough processed image, use 3x3 convolution and ReLU activation function to extract features of the rough processed image again, create a second deep convolution processing chain, wherein the second deep convolution processing chain consists of 5 deep separable convolutions, refine the extracted features, and obtain the output image after integration.

[0054] The spliced ​​crack image is input, and the feature extraction of the rough processed image is performed again by using 3x3 convolution and ReLU activation function, and then more detailed features are extracted through 5 depth-separable layers, and finally the image is output through 3x3 convolution.

[0055] The training layer 13 is used to train the rough processing network and the crack removal network, and to create training sets for the rough processing network and the crack removal network respectively, and to perform training.

[0056] The rough processing network is trained using the training set corresponding to it; and the crack removal network is trained using the training set corresponding to it.

[0057] Figure 3 The block diagram of the structure of the ultra-low computational complexity image enhancement neural network provided by the embodiment of the present invention is shown, and the rough processing network 11 includes:

[0058] The segmentation module 111 is used to segment the input image into a plurality of sub-images.

[0059] The input image is divided into a number of sub-images, and the specific number of sub-images is not limited. Figure 1 Take 2×2 as an example.

[0060] The feature extraction module 112 is used to construct a 3x3 convolution and ReLU activation function, transfer the sub-graph into the 3x3 convolution, extract the features of the sub-graph, and generate a Figure 1 A corresponding feature map, and embed the ReLU activation function into the feature map.

[0061] Select a 3x3 area on the input image, multiply the sub-image by the 3x3 convolution kernel element-by-element, and sum them to produce a single eigenvalue; this process slides on the input image (the unit distance of sliding is also the "step"), and will output a feature map composed of multiple eigenvalues.

[0062] In order to find the nonlinear relationship in the feature map, the output of the 3x3 convolution is transferred to the ReLU activation function to introduce nonlinearity. The benefit of this is that it helps the neural network to better learn and represent complex data relationships.

[0063] Figure 3The structure block diagram of the ultra-low computational complexity image enhancement neural network provided by the embodiment of the present invention is shown, and the rough processing network 11 also includes:

[0064] The feature refinement module 113 is used to create a first deep convolution processing chain, wherein the first deep convolution processing chain is composed of 15 depthwise separable convolutions, transfer the sub-image into the deep convolution processing chain, extract the details of the sub-image, and superimpose the details into the corresponding feature map.

[0065] A first deep convolution processing chain is created, where the first deep convolution processing chain differs from the second deep convolution processing chain only in the number of convolutions. The first deep convolution processing chain is used to extract details in the sub-image and further improve the features.

[0066] The stitching module 114 is used to stitch all the sub-images to obtain a roughly processed image.

[0067] Figure 4 The structure block diagram of the ultra-low computational complexity image enhancement neural network provided by an embodiment of the present invention is shown. The crack removal network 12 includes:

[0068] The re-extraction module 121 is used to perform feature extraction on the roughly processed image again by using 3x3 convolution and ReLU activation function.

[0069] The feature extraction is performed again on the roughly processed image.

[0070] The elimination module 122 is used to create a second depth convolution processing chain, wherein the second depth convolution processing chain consists of 5 depthwise separable convolutions to refine the extracted features.

[0071] The features in the coarsely processed image are transferred to the second deep convolution processing chain to extract the details in the features.

[0072] The output module 123 is used to obtain an output image after integration.

[0073] Features and details are integrated to determine the output image.

[0074] Figure 5 The structure block diagram of the ultra-low computational complexity image enhancement neural network provided by an embodiment of the present invention is shown, and the training layer 13 includes:

[0075] A creation module 131, used to create training sets for the rough processing network and the crack removal network respectively;

[0076] The training module 132 is used for performing training.

[0077] The rough processing network and crack removal network are trained separately using the training set.

[0078] Figure 6 The structure block diagram of the ultra-low computational complexity image enhancement neural network provided by an embodiment of the present invention is shown, and the creation module 131 includes:

[0079] The definition unit 1311 is used to use the 91Images data set in the existing data as the original data, perform grayscale conversion on the original data to obtain a grayscale image, and extract 25 image blocks from each of the grayscale images with a step length of 20, which are defined as clean image blocks.

[0080] The adding unit 1312 is used to add salt and pepper noise of random density to the clean image block to obtain a noisy image block.

[0081] The generating unit 1313 is used to establish a mapping between the noisy image block and the clean image block, generate a mapping pair, summarize all the mapping pairs, obtain a training set corresponding to the rough processing network, and define it as DtasetA.

[0082] The method for constructing the mapping pair of noise image and clean image is as shown in the attached manual. Figure 7 As shown in the figure; the 91Images dataset is used as the original data, and the color image is converted to black and white and the resolution is adjusted. With a step size of 20, 25 image blocks are cut out from each image as clean image blocks; random density salt and pepper noise is added to these image blocks to obtain noisy image blocks. Finally, the noisy image blocks and the clean image blocks are combined together to form a one-to-one corresponding noise image block-clean image block; wherein, the set composed of the noise image block-clean image block is DtasetA.

[0083] The specific process of using DtasetA to rough process the network is as follows: calling keras on the Tensorflow platform, setting the training batch to 8, the initial learning rate to 0.001 and configuring the learning rate decay mechanism, using the RTX3080 GPU, importing DtasetA, carrying out machine training, and training for 100 cycles to reach convergence.

[0084] Through calculation, the computational complexity of the coarse processing network is 93.42 MFLOPS; however, since the principle of the coarse processing network is to divide the image into multiple sub-images and train them separately, and then merge the trained sub-images, this will inevitably lead to splicing cracks at the splicing of the sub-images; therefore, the splicing cracks need to be removed.

[0085] Figure 8 The structure diagram of the ultra-low computational complexity image enhancement neural network provided by the embodiment of the present invention is shown, and the creation module 131 also includes:

[0086] A selection unit 1314 is used to select Pascal VOC from the existing data set, wherein the total number of images of Pascal VOC is 17125, convert the Pascal VOC into a grayscale image, and number the image;

[0087] The inserting unit 1315 is used to insert salt and pepper noise of random density into the numbered Pascal VOC grayscale image to obtain a training set corresponding to the crack removal network, and define it as DatasetB.

[0088] The 17125 color images in Pascal VOC are converted into grayscale images and numbered from 0 to 17124. After adding random density salt and pepper noise, they are processed using the trained rough processing network to obtain the corresponding rough processing images.

[0089] The specific process of using DatasetB for the crack removal network is as follows: calling keras on the Tensorflow platform, setting the training batch to 8, the initial learning rate to 0.001 and configuring the learning rate decay mechanism, using the RTX3080 GPU, importing DatasetB, carrying out machine training, and training for 100 cycles to reach convergence.

[0090] The structural block diagram of the ultra-low computational complexity image enhancement neural network provided by the embodiment of the present invention, wherein the training layer 13 further includes:

[0091] The evaluation module 133 is used to evaluate the image enhancement effects of the rough processing network and the crack removal network.

[0092] By constructing an evaluation module, the processing results of the rough processing network and the crack removal network can be evaluated, thereby providing decision support for the adjustment of the neural network.

[0093] Figure 8 The block diagram of the structure of the ultra-low computational complexity image enhancement neural network provided by the embodiment of the present invention is shown, and the evaluation module 133 includes:

[0094] The construction unit 1331 is used to select the BSD300 image set and add salt and pepper noise of random density to construct an evaluation image set.

[0095] The output unit 1332 is used to determine the processing order of the evaluation image set, and input the evaluation image set into the rough processing network and the crack removal network in sequence according to the processing order to output the enhanced image.

[0096] The BSD300 image set is added with salt and pepper noise of random density and transferred to the rough processing network and the crack removal network in turn. The signal-to-noise ratio is evaluated by PSNR, the information loss is evaluated by MSE, and the computational complexity is evaluated by MFLOPS. Specifically, the larger the PSNR, the better, the smaller the MSE, the better, and the smaller the MFLOPS, the better.

[0097] Figure 8 The block diagram of the structure of the ultra-low computational complexity image enhancement neural network provided by the embodiment of the present invention is shown, and the evaluation module 133 includes:

[0098] The result unit is used to evaluate the signal-to-noise ratio using PSNR, evaluate the information loss using MSE, evaluate the computational complexity using MFLOPS, generate an evaluation table, and use the evaluation table to correct the rough processing network and the crack removal network.

[0099] An evaluation table is generated using the evaluation results, and the rough processing network and the crack removal network are corrected using the evaluation table.

[0100] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0101] The functions that can be realized by the ultra-low computational complexity image enhancement neural network are all completed by a computer device, which includes one or more processors and one or more memories, wherein at least one program code is stored in the one or more memories, and the program code is loaded and executed by the one or more processors to realize the functions of the ultra-low computational complexity image enhancement neural network.

[0102] The processor takes out instructions from the memory one by one, analyzes the instructions, and then completes the corresponding operations according to the instruction requirements, generating a series of control commands, so that the various parts of the computer can automatically, continuously and coordinately move to become an organic whole, realize the input of programs, the input of data, and the calculation and output of results. The arithmetic operations or logical operations generated in this process are all completed by the operator; the memory includes a read-only memory, which is used to store computer programs, and a protection device is provided outside the memory.

[0103] The processor may be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the terminal device, and uses various interfaces and lines to connect various parts of the entire user terminal.

[0104] The above-mentioned memory can be used to store computer programs and / or modules. The above-mentioned processor realizes various functions of the above-mentioned terminal device by running or executing the computer programs and / or modules stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as an information collection template display function, a product information release function, etc.), etc.; the data storage area can store data created according to the use of the berth status display system (such as product information collection templates corresponding to different product types, product information that different product providers need to release, etc.), etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card, a secure digital card, a flash memory card, at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0105] If the module / unit integrated in the terminal device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the modules / units in the above-mentioned embodiment system, and can also be completed by instructing the relevant hardware through a computer program. The above-mentioned computer program can be stored in a computer-readable storage medium, and the computer program can realize the functions of the above-mentioned various system embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. Computer-readable media may include: any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory, random access memory, electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0106] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0107] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An ultra-low computational complexity image enhancement neural network, characterized in that: include: Rough processing network, crack removal network and training layer; The coarse processing network is used to perform preliminary image enhancement on the input image, divide the input image into a plurality of sub-images, construct a 3x3 convolution and a ReLU activation function, transfer the sub-image into the 3x3 convolution, extract features in the sub-image, generate a feature map corresponding to the sub-image one by one, and embed the ReLU activation function into the feature map, create a first deep convolution processing chain, wherein the first deep convolution processing chain consists of 15 deep separable convolutions, transfer the sub-image into the deep convolution processing chain, extract details of the sub-image, and superimpose the details into the corresponding feature map, and splice all the sub-images to obtain a coarse processing image; The crack removal network is used to eliminate the splicing cracks in the rough processed image, use 3x3 convolution and ReLU activation function to extract features of the rough processed image again, create a second deep convolution processing chain, wherein the second deep convolution processing chain consists of 5 depthwise separable convolutions, refine the extracted features, and obtain an output image after integration; The training layer is used to train the rough processing network and the crack removal network, and to create training sets for the rough processing network and the crack removal network respectively, and to perform training.

2. The ultra-low computational complexity image enhancement neural network according to claim 1, characterized in that: The coarse processing network comprises: A segmentation module is used to segment the input image into several sub-images; The feature extraction module is used to construct a 3x3 convolution and a ReLU activation function, transfer the sub-image into the 3x3 convolution, extract the features in the sub-image, generate a feature map corresponding to the sub-image one by one, and embed the ReLU activation function into the feature map.

3. The ultra-low computational complexity image enhancement neural network according to claim 2, characterized in that: The coarse processing network also includes: A feature refinement module, configured to create a first deep convolution processing chain, wherein the first deep convolution processing chain is composed of 15 depthwise separable convolutions, transfer the sub-image into the deep convolution processing chain, extract the details of the sub-image, and superimpose the details into the corresponding feature map; The stitching module is used to stitch all the sub-images to obtain a roughly processed image.

4. The ultra-low computational complexity image enhancement neural network according to claim 1, characterized in that: The crack removal network comprises: A re-extraction module, used for re-extracting features from the roughly processed image using a 3x3 convolution and a ReLU activation function; An elimination module, configured to create a second deep convolution processing chain, wherein the second deep convolution processing chain consists of five depthwise separable convolutions to refine the extracted features; The output module is used to obtain the output image after integration.

5. The ultra-low computational complexity image enhancement neural network according to claim 1, characterized in that: The training layer includes: A creation module, used to create training sets for the rough processing network and the crack removal network respectively; Training module, used for training.

6. The ultra-low computational complexity image enhancement neural network according to claim 5, characterized in that: The creation module includes: A definition unit is used to use the 91Images dataset in the existing data as original data, perform grayscale conversion on the original data to obtain a grayscale image, and extract 25 image blocks from each of the grayscale images with a step length of 20, which are defined as clean image blocks; An adding unit, used for adding salt and pepper noise of random density to the clean image block to obtain a noisy image block; The generation unit is used to establish a mapping between the noisy image block and the clean image block, generate a mapping pair, summarize all the mapping pairs, obtain a training set corresponding to the rough processing network, and define it as DtasetA.

7. The ultra-low computational complexity image enhancement neural network according to claim 5, characterized in that: The creation module also includes: A selection unit is used to select PascalVOC from an existing data set, wherein the total number of images of Pascal VOC is 17125, convert the PascalVOC into a grayscale image, and number the images; The insertion unit is used to insert random density salt and pepper noise into the numbered PascalVOC grayscale image to obtain the training set corresponding to the crack removal network and define it as DatasetB.

8. The ultra-low computational complexity image enhancement neural network according to claim 1, characterized in that: The training layer also includes: The evaluation module is used to evaluate the image enhancement effect of the rough processing network and the crack removal network.

9. The ultra-low computational complexity image enhancement neural network according to claim 8, characterized in that: The evaluation module includes: A construction unit is used to select the BSD300 image set and add salt and pepper noise of random density to construct an evaluation image set; The output unit is used to determine the processing order of the evaluation image set, and input the evaluation image set into the rough processing network and the crack removal network in sequence according to the processing order to output the enhanced image.

10. The ultra-low computational complexity image enhancement neural network according to claim 9, characterized in that: The evaluation module also includes: The result unit is used to evaluate the signal-to-noise ratio using PSNR, evaluate the information loss using MSE, evaluate the computational complexity using MFLOPS, generate an evaluation table, and use the evaluation table to correct the rough processing network and the crack removal network.

Images