LU505937B1 - MASKRCNN WATER SEEPAGE DETECTION METHOD AND SYSTEM BASED ON LOW-LIGHT COMPENSATION - Google Patents

Abstract

Disclosed is a MaskRCNN water seepage detection method and system based on low-light compensation, which includes: S1, extending a sample dataset using fused sample data enhancement; S2, performing data labeling on the extended dataset using Lableme and generating a labeled file of a water accumulated region; S3, performing an enhancement operation on the labeled dataset; S4, training a MaskRCNN model using the enhanced dataset; S5, importing the image to be detected into an MBLLEN model and obtaining a final image of a water seepage region enhanced by a low light; and S6, importing the final output water seepage image enhanced by the low light into the MaskRCNN model, obtaining a feature map by convolution calculation on the water seepage image, obtaining a region proposal through RPN, and obtaining a final position of the water seepage region through an ROI Align layer. A low-light enhancement is performed by importing an inspection image to be detected into an MBLLEN model, and then the low-light enhanced image is imported into a MaskRCNN model for water-accumulated region detection, thus not only performing effective object detection, but also accurately segmenting of an object region.

Description

Ref: PCT/CN2022/134451

MASKRCNN WATER SEEPAGE DETECTION METHOD AND SYSTEM

BASED ON LOW-LIGHT COMPENSATION

Field of Invention

The present disclosure relates to the technical field of identification algorithm optimization of inspection robots in complex indoor environments, and in particular, to a MaskRCNN water seepage detection method and system based on low-light compensation.

Background of the Invention

During operation, hydraulic turbine units often have frequent leakage of the spindle seal, which seriously affects the stable operation of the units. Accidents such as short circuits are more likely for hydraulic turbine layers where cables are distributed. When the leakage is large, there is a serious hidden danger of water accumulation in the hydraulic turbine layers. At the same time, the equipment failure caused by the dripping and leaking of equipment in the hydraulic turbine layers should be timely overhauled to maintain the stable operation of production. Therefore, it will effectively improve the safety factor and reduce the economic loss and potential safety hazards caused by water seepage and leakage through regular and comprehensive dripping and leaking inspection of the inspection region of the equipment in the hydraulic turbine layers, analysis of the overall water seepage situation, timely repair and maintenance according to the water seepage form and degree thereof. However, due to poor light conditions in the hydraulic turbine layers, the images taken by the inspection robots make it difficult to distinguish the boundaries of the water seepage and leakage regions even in the case of light supplements.

Summary of the Invention 1

In order to solve the technical problems in the prior art, the present disclosure LU505937 provides a MaskRCNN water seepage detection method and system based on low-light compensation. A low-light enhancement is performed by importing an inspection image to be detected into an MBLLEN model, and then the low-light enhanced image is imported into a MaskRCNN model for water-accumulated region detection, thus not only performing effective object detection, but also accurately segmenting of an object region.

The present disclosure is achieved by the following technical solutions: a

MaskRCNN water seepage detection method based on low-light compensation, including:

S1, capturing an accumulated water image, collecting a seepage image for accumulated water on the network, and extending a sample dataset using fused sample data enhancement;

S2, performing data labeling on the extended dataset using Lableme and generating a labeled file of a water-accumulated region;

S3, performing an enhancement operation on the labeled dataset, performing operations of flipping, scaling, and gamut changing on the image, and restoring the image to a pixel size of an original image upon completion of the operations;

S4, training a MaskRCNN model using the enhanced dataset;

S5, importing the image to be detected into an MBLLEN model, obtaining a feature map of each layer through layers of a feature extraction module FEM, obtaining an image of each feature map enhanced by a low light through layers of an enhancement module EM, inputting the enhanced feature maps into layers of a fusion module FM, and obtaining a final water seepage region image enhanced by the low light; and

S6, importing the final output water seepage image enhanced by the low light into the MaskRCNN model, obtaining a feature map by convolution calculation on the 2 water seepage image, obtaining a region proposal through RPN, and obtaining a final LUS05937 water seepage region position through an ROI Align layer.

The system of the present disclosure is achieved by the following technical solutions: a MaskRCNN water seepage detection system based on low-light compensation, including: a fused sample data enhancement module configured to extend a sample dataset with a captured accumulated water image and a water seepage image for accumulated water collected on the network using fused sample data enhancement; a data labeling module configured to perform data labeling on the extended dataset using Lableme and generate a labeled file of a water-accumulated region; an enhancement operation module configured to perform an enhancement operation on the labeled dataset, perform operations of flipping, scaling, and gamut changing on the image, and restore the image to a pixel size of an original image upon completion of the operations; a MaskRCNN model training module configured to train a MaskRCNN model using the enhanced dataset; a low-light enhanced water seepage region image acquisition module configured to import the image to be detected into an MBLLEN model, obtain a feature map of each layer through layers of a feature extraction module FEM, obtain an image of each feature map enhanced by a low light through layers of an enhancement module

EM, input the enhanced feature maps into layers of a fusion module FM, and obtain a final water seepage region image enhanced by the low light; and a water seepage region position acquisition module configured to import the final output water seepage image enhanced by the low light into the MaskRCNN model, obtain a feature map by convolution calculation on the water seepage image, obtain a region proposal through RPN, and obtain a final water seepage region position through an ROI Align layer. 3

The present disclosure has the following advantages and beneficial effects compared to the prior art.

A low-light enhancement is performed by importing an inspection image to be detected into an MBLLEN model, and then the low-light enhanced image is imported into a MaskRCNN model for water-accumulated region detection, thus not only performing effective object detection, but also accurately segmenting boundaries of an object region.

Brief Description of the Drawings

Fig. 1 is a flowchart of a method of the present disclosure;

Fig. 2 (a) is a schematic diagram for capturing an accumulated water image according to the present disclosure;

Fig. 2 (b) is a schematic diagram of a water seepage image for accumulated water collected on the network;

Fig. 2 (c) is a schematic diagram of fused sample data enhancement according to the present disclosure;

Fig. 3 is a schematic diagram of an image of a water-accumulated region labeled by the Labelme software;

Fig. 4 is a structural schematic diagram of a MaskRCNN model;

Fig. 5 is a structural schematic diagram of an MBLLEN model;

Fig. 6 (a) is a low-light image; and

Fig. 6 (b) is an enhanced image. 4

Detailed Description of the Embodiments LUS05937

Hereinafter, the present disclosure will be explained in detail with reference to the embodiments and the accompanying drawings, but the embodiments of the present disclosure are not limited thereto.

Embodiment

As shown in Fig. 1, the MaskRCNN water seepage detection method based on low-light compensation in the present embodiment includes the following steps.

At S1, an accumulated water image is captured, a similar water seepage image for accumulated water is collected on the network, and a sample dataset is expanded using fused sample data enhancement.

At S2, data labeling is performed on the extended dataset using Lableme and a labeled file of a water-accumulated region is generated.

At S3, an enhancement operation is performed on the labeled dataset, operations of flipping, scaling, gamut changing, and the like are performed on the image, and the image is restored to a pixel size of an original image upon completion of the operations.

At S4, a MaskRCNN model is trained using the enhanced dataset.

At S5, the image to be detected is imported into an MBLLEN model, a feature map of each layer is obtained through layers of a feature extraction module FEM, an image of each feature map enhanced by a low light is obtained through layers of an enhancement module EM, the enhanced feature maps are input into layers of a fusion module FM, and a final water seepage region image enhanced by the low light is obtained.

At S6, the final output water seepage image enhanced by the low light is imported into the MaskRCNN model, a feature map is obtained by convolution 5 calculation on the water seepage image, a region proposal is obtained through RPN, LUS05937 and a final water seepage region position is obtained through an ROI Align layer.

As shown in Fig. 2 (a), Fig. 2 (b), and Fig. 2 (c), in this embodiment, a specific process of the fused sample data enhancement in S1 is as follows.

Two images are randomly selected from a training set and are enhanced using a data enhancement method including flipping, adding noise, cropping, and the like, and the two accumulated water images are fused according to random weights to increase sample diversity; Specifically, a formula for fusing the two images is as follows:

Image(R,G, B) = nxImagel(R,G, B) + (1-177) Image2(R,G, B) 7 = rand(0.3—0.7) (1) where Image(R.G,B) ;the fused image, Imagel(R,G,B) and

Image2(R,G,B) are the original two images, 7 = rand(0.3—0.7) represents random numbers with a fusion weight of 0.3 to 0.7, and R,G,B are three channels of the images.

As shown in Fig. 3, in this embodiment, a specific process of the data labeling in

S2 is as follows. the water seepage region is labelled by a multi-line and multi-point method; a polygonal contour of the water seepage region is labelled using a labeling tool

Accordingly, a label name of the water seepage contour is set, a corresponding json file to the labeled sample is generated, and contour and image information of the object region in the sample is stored by the json file.

As shown in Fig. 4, in this embodiment, the MaskRCNN is an example segmentation framework for performing effective object detection and accurately segmenting boundaries of the object region. The MaskRCNN model mainly includes a feature extraction framework ResNet and an RPN module; the ResNet extracts features of the image to be detected using a multi-layer convolution structure, and the 6

RPN is configured to generate a plurality of ROI regions; the MaskRCNN replaces LUS05937

Rol Pooling using an Rol Align layer, and maps the plurality of ROI feature regions generated by the RPN to a uniform size of 7*7 using bilinear interpolation; Finally, the plurality of ROI regions generated by the RPN layer are classified and a regression operation of a positioning box is performed on the same, and a Mask corresponding to the water seepage region is generated using a full convolution neural network FCN.

In this embodiment, a loss function Loss of the MaskRCNN is defined as:

Loss = L, + L,, + L, 2) where ““ is a classification error, x is an error generated by the positioning box, and Last is an error caused by the Mask; the classification error La is constructed by introducing a log-likelihood loss, and a calculation formula of La is as follows: 1 N M

L,=-log P(Y|X)=-— X v,log(p;)

Nine (3) where X and Y are a testing classification and an actual classification, respectively, N is the number of input samples, M is the number of possible classes, and Pi represents the probability distribution of class j predicted and output by a model of sample xi; Yi represents whether the true class of the sample xi is class j; in order to increase the robustness of the loss function, the error Lox caused by the positioning box adopts LI loss; and the relative entropy of pixels in the ROI region is calculated using a sigmoid function and the average relative entropy error Loos is obtained.

In order to get better generalization performance on the MaskRCNN for a small 7 labeled dataset, this embodiment introduces a pre-trained weight (mask_renn_coco.h5) LUS05937 on a COCO dataset for fine tuning. MBLLEN and MaskRCNN detection models of water-accumulated regions are used for classification. The inspection image to be detected is imported into the MBLLEN model for low-light enhancement, then the low-light enhanced image is imported into the MaskRCNN model for water-accumulated region detection, and the labeled water seepage region is output.

As shown in Fig. 5, in this embodiment, the MBLLEN model is a multi-layer low-light enhancement depth learning network model. Image features of different layers are extracted by convolution calculation, and feature maps of different layers are input into a plurality of sub-networks for enhancement. The MBLLEN model mainly includes a feature extraction module (FEM), an enhancement module (EM), and a fusion module (FM). The feature extraction module FEM includes 10 layers of unidirectional network structure, where each layer adopts 32 convolution kernels of a size of 3x3, has a stride of 1, and an activation function thereof is ReLU. The feature extraction module does not adopt a pooling layer. The output of each layer is simultaneously the input of a convolution layer of the next feature extraction module

FEM and the input of a corresponding convolution layer of the enhancement module

EM. Since the feature extraction module FEM includes 10 feature extraction layers, the enhancement module EM includes 10 sub-network structures with the same structure. The sub-network structures of the EM include a convolution layer, 3 convolution layers, and 3 deconvolution layers. The fusion module FM fuses all images output from the sub-networks of the EM and obtains the final enhancement result by a 3-channel convolution kernel of a size of 1x1.

In order to train the MBLLEN model for image low-light compensation, a structure loss (Str), a pre-training VGG content loss (VGG), and a region loss are defined respectively.

Specifically, the formula of the loss function is as follows:

Loss = Lg, + Lyon, + Lresion (4) 8 where the structure loss is mainly configured to reduce the structure distortion of LU505937 an enhanced image and a real image, and the specific formula is as follows:

Lg, = Loss + us-ssma (5) where Lions is a structural similarity of the enhanced image and the real image and Loss sm is a multi-level structural similarity; the pre-training VGG content loss minimizes the absolute difference between the enhanced image and the real image output in the pre-training VGG-19 network, and the formula of the loss function is as follows: 1 W, H,; Cj

Lycon.; = Te Le AE), 7 Pi (G)…] i,j rij x=1 y=l z=1 (6) where E and G are the enhanced image and the real image, respectively, Weg ,

H, , and Gus represent dimensions of the feature map of the pre-training VGG, respectively; bi represents the j convolution layer and ith feature map of the

VGG-19 network; x, y, and z represent the width, height, and channel number of the feature map, respectively; the region loss approximates a dark region of the whole image by segmenting 40% darkest pixel value of the image, and obtains the following loss function: +R , , 1 te , ,

Legon F5 DONNE, D) = Go pl) + wy — ED Gui, D)

MN, ia j= MR, 5 A (7) where E and G, are low-light regions of the enhanced image and the real image, respectively, Ey and Gu are non-low-light regions of the enhanced image and the real image, respectively, * and ” are 4 and 1, respectively; " is the 9 width of the image Gi, "is the height of the image Gi, Mu js the width of the 905987 image Gu :and ”# js the height of the image Gy .

A low-light dataset is synthesized and obtained based on a PASCAL VOC dataset, and Gamma correction and Poisson noise with a peak value of 200 are added as low-light input images, and original images are added as real images. The enhanced image results for low-light water seepage images are shown in Fig. 6 (a) and Fig. 6 (b).

Based on the same inventive concept, the present disclosure proposes a

MaskRCNN water seepage detection system based on low-light compensation, including: a fused sample data enhancement module configured to extend a sample dataset with a captured accumulated water image and a water seepage image for accumulated water collected on the network using fused sample data enhancement;a data labeling module configured to perform data labeling on the extended dataset using Lableme and generate a labeled file of a water-accumulated region; an enhancement operation module configured to perform an enhancement operation on the labeled dataset, perform operations of flipping, scaling, and gamut changing on the image, and restore the image to a pixel size of an original image upon completion of the operations; a MaskRCNN model training module configured to train a MaskRCNN model using the enhanced dataset; a low-light enhanced water seepage region image acquisition module configured to import the image to be detected into an MBLLEN model, obtain a feature map of each layer through layers of a feature extraction module FEM, obtain an image of each feature map enhanced by a low light through layers of an enhancement module

EM, input the enhanced feature maps into layers of a fusion module FM, and obtain a 10 final water seepage region image enhanced by the low light; and LUS05937 a water seepage region position acquisition module configured to import the final output water seepage image enhanced by the low light into the MaskRCNN model, obtain a feature map by convolution calculation on the water seepage image, obtain a region proposal through RPN, and obtain a final water seepage region position through an ROI Align layer.

While the above embodiments are preferred embodiments of the present disclosure, the present disclosure is not limited thereto. Other changes, modifications, substitutions, combinations and simplifications are all equivalent and can be made without departing from the spirit and principles of the present disclosure. 11

Claims (9)

Claims LU505937 I. A MaskRCNN water seepage detection method based on low-light compensation, comprising: S1, capturing an accumulated water image, collecting a water seepage image for accumulated water on the network, and extending a sample dataset using fused sample data enhancement; S2, performing data labeling on the extended dataset using Lableme and generating a labeled file of a water-accumulated region; S3, performing an enhancement operation on the labeled dataset, performing operations of flipping, scaling, and gamut changing on the image, and restoring the image to a pixel size of an original image upon completion of the operations; S4, training a MaskRCNN model using the enhanced dataset; S5, importing the image to be detected into an MBLLEN model, obtaining a feature map of each layer through layers of a feature extraction module FEM, obtaining an image of each feature map enhanced by a low light through layers of an enhancement module EM, inputting the enhanced feature maps into layers of a fusion module FM, and obtaining a final water seepage region image enhanced by the low light; and S6, importing the final output water seepage image enhanced by the low light into the MaskRCNN model, obtaining a feature map by convolution calculation on the water seepage image, obtaining a region proposal through RPN, and obtaining a final water seepage region position through an ROI Align layer. 2. The MaskRCNN water seepage detection method based on low-light compensation according to claim 1, wherein the fused sample data enhancement in S1 specifically comprises: randomly selecting two images from a training set and enhancing the same using a data enhancement method comprising flipping, adding noise, and cropping, and fusing the two accumulated water images according to random weights to increase sample diversity; a formula for fusing the two images is as follows: Im age RG Be In age HE GA + 0-0 im aged(R OF i= renee I~ OT (1) 12 wherein 1mage(R.G.B) io the fused image, Imagel( R,G,B) and 7505957 Image2(R,G,B) are the original two images, 7=rand(0.3=0.7) represents random numbers with a fusion weight of 0.3 to 0.7, and R.G.B are three channels of the image . 3. The MaskRCNN water seepage detection method based on low-light compensation according to claim 1, wherein the data labeling in S2 specifically comprises: labeling the water seepage region by a multi-line and multi-point method; labeling a polygonal contour of the water seepage region using a labeling tool Lableme, setting a label name of the water seepage contour, generating a corresponding json file to the labeled sample, and storing contour and image information of the object region in the sample by the json file. 4. The MaskRCNN water seepage detection method based on low-light compensation according to claim 1, wherein in S4, the MaskRCNN model comprises a feature extraction framework ResNet and an RPN module; the ResNet extracts features of the image to be detected using a multi-layer convolution structure, and the RPN is configured to generate a plurality of ROI regions; MaskRCNN replaces Rol Pooling using an Rol Align layer, and maps the plurality of ROI feature regions generated by the RPN to a uniform size of 7*7 using bilinear interpolation; Finally, the plurality of ROI regions generated by the RPN layer are classified and a regression operation of a positioning box is performed on the same, and a Mask corresponding to the water seepage region is generated using a full convolution neural network FCN. 5. The MaskRCNN water seepage detection method based on low-light compensation according to claim 4, wherein a loss function Loss of the MaskRCNN is defined as: Loss=L, +L, +L, (2) wherein La, is a classification error, Le is an error generated by the positioning box, and Task is an error caused by the Mask; 13 L LU505937 the classification error “«“ is constructed by introducing a log-likelihood loss, and a calculation formula of La, is as follows: 1 N M Ly=-log P(Y]|X)=-—=3%"y,log(p,) Nine (3) wherein X and Y are a testing classification and a real classification, respectively, N is the number of input samples, M is the number of possible classes, and Pi represents the probability distribution of class j predicted and output by a model of sample xi; Yi represents whether the true class of the sample xi is class j; the error Lyon caused by the positioning box adopts L1 loss; the relative entropy of the pixels in the ROI region is calculated using a sigmoid function and the average relative entropy error Task is obtained. 6. The MaskRCNN water seepage detection method based on low-light compensation according to claim 1, wherein the training the MaskRCNN model in S4 comprises performing fine tuning by introducing pre-trained weights on a COCO dataset. 7. The MaskRCNN water seepage detection method based on low-light compensation according to claim 1, wherein implementing the MBLLEN model in S5 specifically comprises: S51, dividing the MBLLEN model into a feature extraction module FEM, an enhancement module EM, and a fusion module FM; S52, the feature extraction module FEM comprising 10 layers of unidirectional network structure, wherein each layer adopts 32 convolution kernels of a size of 3x3, has a stride of 1, and an activation function thereof is ReLU; the output of each layer being simultaneously the input of a convolution layer of the next feature extraction module FEM and the input of a corresponding convolution layer of the enhancement module EM; S53, the enhancement module EM comprising 10 sub-network structures with the same structure comprising a convolution layer, 3 convolution layers, and 3 deconvolution layers; and 14 S54, the fusion module FM fusing all images output from the sub-networks of LU505937 and the enhancement module EM, and obtaining a final enhancement result by a 3-channel convolution kernel of size 1x1. 8. The MaskRCNN water seepage detection method based on low-light compensation according to claim 7, wherein the training the MBLLEN model comprises: defining a structure loss, a pre-training VGG content loss, and a region loss; the formula of the loss function is as follows: Loss = Lg, + Lys; ; + region (4) wherein a specific formula of the structure loss is as follows: Lg, = Loss + us-ssma (5) wherein Lions is a structural similarity between an enhanced image and a real image and Loss sm is a multi-level structural similarity; a formula for the pre-training VGG content loss is as follows: 1 W, H,; Cj Lycon.; = Te Le AE), 7 Pi (G)…] ij Ci x=1 yal = (6) wherein E and G are the enhanced image and the real image, respectively, Wis , H, , and Gus represent dimensions of the feature map of the pre-training VGG respectively; bi represents the j convolution layer and ith feature map of the VGG-19 network; x, y, and z represent the width, height, and channel number of the feature map, respectively; the region loss obtains a dark region of the whole image by segmenting 40% darkest pixel value of the image, and obtains the following loss function: +R , , 1 te , , Ligon TV DE DAL + wy — ED (En (is 7 ) +6, 6. DI) mn i=l j=1 myn, i=l j=1 (7) wherein E and G, are low-light regions of the enhanced image and the real image, respectively, Ey and Gu are non -low-light regions of the enhanced image and the real image, respectively, “i and "# are 4 and 1, respectively; ML js the 15 width of the image Gi, "is the height of the image Gi, Mu js the width of the 905987 image Gu :and ”# js the height of the image Gy . 9. A MaskRCNN water seepage detection system based on low-light compensation, comprising: a fused sample data enhancement module configured to extend a sample dataset with a captured accumulated water image and a water seepage image for accumulated water collected on the network using fused sample data enhancement; a data labeling module configured to perform data labeling on the extended dataset using Lableme and generate a labeled file of a water-accumulated region; an enhancement operation module configured to perform an enhancement operation on the labeled dataset, perform operations of flipping, scaling, and gamut changing on the image, and restore the image to a pixel size of an original image upon completion of the operations; a MaskRCNN model training module configured to train a MaskRCNN model using the enhanced dataset; a low-light enhanced water seepage region image acquisition module configured to import the image to be detected into an MBLLEN model, obtain a feature map of each layer through layers of a feature extraction module FEM, obtain an image of each feature map enhanced by a low light through layers of an enhancement module EM, input the enhanced feature maps into layers of a fusion module FM, and obtain a final water seepage region image enhanced by the low light; and a water seepage region position acquisition module configured to import the final output water seepage image enhanced by the low light into the MaskRCNN model, obtain a feature map by convolution calculation on the water seepage image, obtain a region proposal through RPN, and obtain a final water seepage region position through an ROI Align layer. 16 Claims LU505937 1. A MaskRCNN method for detecting water seepage based on low light compensation, characterized by comprising: S1, acquiring an image of accumulated water, collecting a seepage image for accumulated water on the network, and enhancing a sample data set using fused sample data enhancement; S2, performing data labeling on the enhanced data set using labels and generating a labeled file of a water-accumulated region; S3, performing an enhancement operation on the labeled data set, performing operations of mirroring, scaling, and changing color scale on the image, and restoring the image to a pixel size of an original image after completing the operations; S4, training a MaskRCNN model using the enhanced data set; S5, importing the image to be captured into an MBLLEN model, obtaining a feature map of each layer by layering a feature extraction module FEM, obtaining an image of each feature map enhanced by a low light by layering an enhancement module EM, inputting the enhanced feature maps into layers of a fusion module FM, and obtaining a final image of a water seepage region enhanced by the low light; and S6, importing the final seepage image enhanced by the low light into the MaskRCNN model, obtaining a feature map by convolution calculation on the seepage image, obtaining a region proposal by RPN, and obtaining a final position of the seepage region by an ROI alignment layer. 2. The MaskRCNN method for detecting seepage water based on low light compensation according to claim 1, characterized in that the enhancement of the fused sample data in S1 in particular comprises: randomly selecting two images from a training set and enhancing them using a data enhancement method, the data enhancement method comprising flipping, adding noise and cropping, and merging the two accumulated water images according to random weights to increase the sample diversity; where a formula for merging the two images is as follows: Image(R, G, B) =7xImagel(R,G, B)+(1-7)Image2(R, G, B) 17 = rand(0.3—0.7) (1) where Image(R,G,B) is the merged image, Imagel(R,G,B) and Image2(R,G,B) are the two original images, 7 = rand(0.3—0.7) represent random numbers with a merging weight of 0.3 to 0.7, and R,G,B are three channels of the image. 3. The MaskRCNN seepage water detection method based on low light compensation according to claim 1, characterized in that the data labeling in S2 specifically includes: labeling the seepage water region by a multi-line and multi-point method; labeling a polygonal contour of the seepage water region using a labeling tool, namely Lableme; setting a label name of the seepage water contour; generating a corresponding json file for the labeled sample; and storing contour and image information of the object region in the sample by the json file. 4. The MaskRCNN method for detecting seepage water based on low light compensation according to claim 1, characterized in that in S4, the MaskRCNN model comprises a feature extraction framework ResNet and an RPN module; wherein the ResNet extracts features of the image to be detected using a multi-layer convolution structure, the RPN is configured to generate a plurality of ROI regions; wherein MaskRCNN replaces the Rol-Pooling with a Rol-Align layer and maps the plurality of ROI feature regions generated by the RPN layer to a uniform size of 7*7 using bilinear interpolation; finally, the plurality of ROI regions generated by the RPN layer are classified and a regression operation of a positioning box is performed on the same, and a mask corresponding to the seepage water region is generated using a fully convolutional neural network FCN. 5. The MaskRCNN method for detecting water seepage based on low light compensation according to claim 1, characterized in that a loss function Loss of the MaskRCNN is defined as: Loss = La + Ly, + Last (2) where La is a classification error, Lie is an error generated by the positioning box, and Ent is an error caused by the mask; the classification error Las is constructed by introducing a log-likelihood loss, and a calculation formula of La is as follows: Xu LU505937 L,=~logP (Y|X) = -—> > y, log (ps) NS j=1 (3) where X and Ÿ are a test classification and a real classification, respectively, N is the number of input samples, M is the number of possible classes, and Pi represents the probability distribution of class j predicted and output by a model of sample xi; where Yi represents whether the true class of sample xi is class j; where the error Lie caused by the positioning box assumes the L1 loss; the relative entropy of pixels in the ROI region is calculated using a sigmoid function, and the average relative entropy error Last is obtained. 6. MaskRCNN method for detecting water seepage based on low light compensation according to claim 1, characterized in that training the MaskRCNN model in S4 comprises performing fine-tuning by introducing pre-trained weights on a COCO dataset. 7. MaskRCNN method for detecting water seepage based on low light compensation according to claim 1, characterized in that implementing the MBLLEN model in S5 in particular comprises: S51, dividing the MBLLEN model into a feature extraction module FEM, an enhancement module EM and a fusion module FM; S52, the feature extraction module FEM comprises 10 layers of a unidirectional network structure, each layer adopting 32 convolution kernels of size 3 x 3, having a convolution tride of 1 and an activation function thereof being ReLU; wherein the output of each layer is simultaneously the input of a convolution layer of the next feature extraction module FEM and the input of a corresponding convolution layer of the enhancement module EM; S53, the enhancement module EM includes 10 subnetwork structures with the same structure, which includes a 1-layer convolutional layer, a 3-layer convolutional layer, and a 3-layer deconvolutional layer; and S54, the fusion module FM fuses all images output from the subnetworks of the enhancement module EM, and a final enhancement result is obtained by a 3-channel convolution kernel with a size of 1 x 1. 8. MaskRCNN method for detecting seepage water based on low light compensation according to claim 7, characterized in that the Training the MBLLEN model includes: LUS05937 defining a structural loss, a pre-training VGG content loss, and a region loss; where the formula of the loss function is as follows: Loss = Lg, + Lyooh,j + Lpegion (4) where a specific formula for the structural loss is as follows: 1 We; H, Cij Ligon, = WHC >>> |e. (EB), ,. —-P, G),,. ijt ii, xl y=l z=1 (5) where Lg, is a structural similarity between an enhanced image and a real image, and L, son, is a multi-level structural similarity; where a formula for the pre-training VGG content loss is as follows: 1 We; H, Cij Ligon, = WHC >>> |e. (EB), ,. —-P, G),,. ijt ii, xl y=l z=1 (6) where E and G are the enhanced image and the real image, respectively, where W, ;,H, ;and C,, represent the dimensions of the feature map of the VGG before training; where ®,, represents the j-th convolutional layer and the i-th feature map of the VGG-19 network; where x, y and z represent the width, height and channel number of the feature map, respectively; where the region loss obtains a dark region of the whole image by segmenting 40% of the darkest pixel value of the image and obtains the following loss function: 1 GE . , I A , , Ligon = Wom EG DD GL pl) + m ED GG DI) mn, i=l j=1 mans i=l j=1 (7) where E, and G, are low-light regions of the enhanced image and the real image, respectively, E, and G, are non-low-light regions of the enhanced image and the real image, respectively, w, and w, are 4 and 1, respectively; m, is the width of the image G,; n, is the height of the image G,; mn, is the width of the image G,, and n, is the height of the image G,. 9. A MaskRCNN system for detecting water seepage based on low light compensation, characterized by comprising: a fused sample data enhancement module configured to enhance a sample data set including a captured water-accumulated image and a water-accumulated image LUS05937 for accumulated water collected in the network using fused sample data enhancement; a data labeling module configured to perform data labeling on the enhanced data set using Lableme and generate a labeled file of a water-accumulated region; an enhancement operation module configured to perform an enhancement operation on the labeled data set, perform mirroring, scaling, and changing the color gamut on the image, and restore the image to a pixel size of an original image after completion of the operations; a MaskRCNN model training module configured to train a MaskRCNN model using the improved dataset; a seepage region image acquisition module with low-light enhanced, configured to import the image to be acquired into an MBLLEN model, obtain a feature map of each layer by layering a feature extraction module FEM, an image of each feature map enhanced by low-light enhanced by layering an enhancement module EM, input the enhanced feature maps into layers of a fusion module FM, and obtain a final image of the seepage region enhanced by the low-light; and a seepage region position acquisition module configured to import the final output image of the seepage region enhanced by the low-light into the MaskRCNN model, obtain a feature map by convolution calculation on the seepage image, obtain a region proposal by RPN, and obtain a final position of the seepage region by an ROI alignment layer.