Hydropower station water collecting well video monitoring image water mist removing method

Abstract

The invention discloses a hydropower station water collecting well video monitoring image water mist removing method, which comprises the following steps of S1, capturing a water collecting well image through a monitoring camera, wherein the water collecting well image comprises an image set # imgabs0 # with water mist and an image set # imgabs1 # without water mist; s2, performing manual annotation on the image sets # imgabs2 # and # imgabs3 # to make a data set; s3, constructing a water mist removal neural network training model; s4, training a water mist removal neural network training model; s5, performing image water mist removal by using the trained water mist removal model; according to the technical scheme, the artificial intelligence algorithm can be used for removing water mist from the hydropower station water-collecting well video image; the problem that a hydropower station water-collecting well video monitoring system is easily interfered by water mist is effectively solved.

Description

Technical Field

[0001] The invention relates to the technical field of video monitoring of water collection wells in hydropower stations, in particular to a method for removing water mist from video monitoring images of water collection wells in hydropower stations. Background Art

[0002] The sump is an important part of the drainage system of a hydropower station. During the maintenance period of the unit, the remaining water in the volute, gate leakage, seepage in the dam hydraulic structure, equipment leakage and domestic water are all collected in the sump through the drainage ditch and discharged to the downstream by a deep well pump.

[0003] Water level monitoring in water collection wells is an important part of the operation and maintenance of hydropower stations. Once the deep well pump fails or the pumping and drainage efficiency decreases, the water level in the water collection well will continue to rise, which will cause flooding of the power plant. In order to further improve the safety and reliability of the water collection well water level monitoring system, a giant hydropower station has deployed a video monitoring system in the water collection well. By shooting the water gauge, the water level of the water collection well is calculated in real time, and the water level signal collection and danger alarm are realized through linkage with the power station computer monitoring system and SMS alarm.

[0004] However, in actual applications, the seepage flow of hydropower stations is large, and the elevation difference between the confluence and the working water level of the collection well is large, resulting in a large amount of water mist, which interferes with the video surveillance system in collecting water level images. Summary of the invention

[0005] The purpose of the present invention is to overcome the above-mentioned shortcomings and provide a method for removing water mist from a video monitoring image of a water collection well of a hydropower station, thereby solving the problem that the video monitoring system of the water collection well of a hydropower station is easily disturbed by water mist.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for removing water mist from a video monitoring image of a water collection well of a hydropower station, which comprises the following steps: S1. Capture the water collection well image through the surveillance camera, including the image set with water mist and a collection of images without water mist ; S2. Image collection and Perform manual annotation to create data sets; S3, building a neural network training model for water mist removal; S4, training a neural network training model for water mist removal; S5. Use the trained water mist removal model to remove water mist from the image.

[0007] Furthermore, the specific process of step S2 is as follows: Use labelimg to label Annotate the images in the collection, generate candidate frames for the water fog area in the image, and obtain the candidate frames w, h, x, y, where w and h are the relative width and height of the candidate frame, the relative width and height are the width and height of the candidate frame divided by the width and height of the image, and x and y are the relative center coordinates of the candidate frame. The same image contains multiple candidate frames. After marking a After the images in the collection, The image with the same water level as the annotated image is used as the real image after removing the water mist; Mark one by one The images in the collection are searched and the corresponding paired images are found to complete the dataset creation; The data set is randomly divided into a training set and a validation set in a ratio of 8:2.

[0008] Furthermore, the water mist removal neural network training model in step S3 is composed of an input, an encoder-decoder network, and a discriminant network; wherein the encoder-decoder network is used to generate an image after water mist removal from the original input image, which includes an encoder network and a decoder network; the encoder network includes an encoder skeleton network and a void space pyramid pooling module.

[0009] Furthermore, the encoder skeleton network includes: Initialize the convolutional layer, which consists of a convolutional layer with a kernel size k of 7, a stride s of 4, a zero padding method p of [2,1,2,1], and a layer normalization LN. [2,1,2,1] means 2 layers of zero padding on the left and above the image, and 1 layer of zero padding on the bottom and right of the image. 3 ConvNeXt modules with 64 output channels, where the ConvNeXt module contains a residual structure; a depth-wise separable convolutional layer with k=7, s=1, and p=3; layer normalization LN; a convolutional layer with k=1 and s=1; a GELU activation function; a convolutional layer with k=1 and s=1; layer scaling and a Drop Path layer, whose output features are the feature stacking of the Drop Path layer output features and the ConvNext module input features; Downsampling module,The downsampling module consists of layer normalization LN and convolutional layers with k=7, s=2, and p=[3,2,3,2]; 9 ConvNeXt modules with 128 output channels, a downsampling module, and 3 ConvNeXt modules with 256 output channels. The encoder network outputs 240*240*64 low-dimensional features and 60*60*256 high-dimensional features in the 3rd ConvNeXt and 15th ConvNeXt modules respectively; The atrous spatial pyramid pooling module includes: Input, i.e., the high-dimensional features output by the encoder skeleton network; 5 branches; feature concatenation and 1*1 convolution, which includes convolution with k=1 and s=1, layer normalization LN, GELU activation function; 4x bilinear interpolation; Among them, branch 1 is 1*1 convolution; branch 2 is a dilated convolution with k=3, s=1, p=6, and expansion coefficient d=6, and layer-normalized LN and GELU activation functions; branch 3 is a dilated convolution with k=3, s=1, p=12, and d=12, and layer-normalized LN and GELU activation functions; branch 4 is a dilated convolution with k=3, s=1, p=18, and d=18, and layer-normalized LN and GELU activation functions; branch 5 is average pooling, 1*1 convolution, and 2x bilinear interpolation; the features output by the 5 branches are spliced, input into 1*1 convolution and 4x bilinear interpolation; the dilated spatial pyramid pooling module is used to collect image features of different scales, increase the network's receptive field, and enhance the network's ability to extract global image information; The decoder network includes: input, that is, the low-dimensional features output by the encoder skeleton network; 1*1 convolution; feature splicing, that is, splicing the 1*1 convolution output features and the features output by the encoder network; convolution and layer normalization LN and GELU activation functions with k=3, s=1, and p=1; 4x bilinear interpolation.

[0010] Furthermore, the discriminant network is used to assist the encoder-decoder network in training the ability to remove water mist. Its structure includes two branches, branch 1 includes input, that is, the features of the encoder-decoder output; the convolution layer is initialized, which consists of a convolution layer with a convolution kernel size k of 7, a step size s of 4, a zero padding method p of [2,1,2,1] and a layer normalization LN; a ConvNeXt module with an output channel of 128 and a downsampling module; a ConvNeXt module with an output channel of 256 and a downsampling module; branch 2 includes input, that is, the encoder-decoder The image cropped by the four parameters w, h, x, and y of the true values ​​corresponding to the features output by the network; 1*1 convolution; ConvNeXt module with an output channel of 128, downsampling module; ConvNeXt module and downsampling module with an output channel of 256; feature splicing of the output features of branch 1 and branch 2, input 1*1 convolution, ConvNeXt module with an output channel of 512, downsampling module, ConvNeXt module with an output channel of 1024, downsampling module, global average pooling and softmax layer.

[0011] Furthermore, when constructing the water mist removal neural network training model in step S3, the water mist RGB image with a size of 960*960 is used as the input of the model, and an image without water mist is generated through the encoder-decoder network. The corresponding manually marked candidate box area is cropped and scaled to a size of 240*240, and the generated image without water mist and the cropped and scaled image are used as input, and the discriminant network calculates and outputs the confidence of the generated image without water mist.

[0012] Furthermore, the specific process of step S4 is as follows: 4.1. Scale the training set data to 960*960 and normalize it. Crop and scale the manually annotated area in the training data set to 240*240 and normalize it. 4.2. Set the basic training parameter BATCH_SIZE to 8; 4.3. When the number of training rounds EPOCH is 200 or less, only the encoder-decoder network is trained, Adam is the training optimizer, the initial learning rate is 0.01, and its loss function is set to mean square error loss , whose formula is 2 ;in Generate the pixel value of the i-th pixel for the input image annotation area through neural network prediction, is the true pixel value of the i-th pixel, and stores the parameter P1 that performs best in the validation set at this stage; 4.4. When the training round number EPOCH∈(200,500], the training model is trained as a whole, the encoder-decoder network loads the parameter P1, Adam is used as the training optimizer, the initial learning rate is 0.001, and its loss function is set to + ,in , where D is the discriminator, i.e., the discriminant network, G is the generator, i.e., the encoder-decoder network, x is the real image, G(z) is the image generated by the encoder-decoder network, and the parameter P2 with the best performance on the validation set at this stage is stored.

[0013] Furthermore, the specific process of step S5 is as follows: 5.1. Obtain the water level image captured by the water collection well monitoring camera; 5.2. Scale the image to 960*960 size; 5.3. Normalize the scaled image; 5.4. Use the encoder-decoder network to load the corresponding parameters in P2; 5.5. Input the normalized image into the encoder-decoder network, which outputs the dehazed image; 5.6. Scale the dehazed image to its original size to complete the dehazing process.

[0014] Beneficial effects of the present invention: (1) The technical solution of the present invention can use artificial intelligence algorithms to remove water mist from video images of water collection wells in hydropower stations, effectively solving the problem that the video monitoring system of water collection wells in hydropower stations is easily disturbed by water mist.

[0015] (2) The water mist removal neural network model described in the scheme of the present invention has certain advancements in model design. Its encoder-decoder network fully integrates features of multiple dimensions for water mist removal, and is trained in combination with the idea of ​​generative adversarial networks to further improve the accuracy of water mist removal. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a schematic diagram of a water collection well; Figure 2 This is a schematic diagram of the water mist removal neural network training model structure; Figure 3 This is the encoder-decoder network structure diagram; Figure 4 This is the encoder skeleton network structure diagram; Figure 5 This is the structure diagram of the ConvNeXt module and the downsampling module. DETAILED DESCRIPTION

[0017] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0018] Example 1: Application scenario of the present invention is as follows Figure 1 As shown in the figure, 1 is a monitoring camera, which is deployed at the mouth of the water collection well; 2 is the confluence of leakage water, which is discharged into the water collection well by a pipe; 3 is a deep well pump, which discharges the water in the water collection well to the downstream of the dam when working; 4 is a submersible sewage pump, which is used to discharge the dirt at the bottom of the water collection well; 5 is a fixed water gauge, which cooperates with the monitoring camera to calculate the water level of the water collection well.

[0019] A method for removing water mist from a video monitoring image of a water collection well in a hydropower station comprises the following steps: S1. Capture the water collection well image through the surveillance camera, including the image set with water mist and a collection of images without water mist ; S2. Image collection and Manually label and create a data set: Use labelimg labeling software to The images in the collection are annotated to generate candidate frames of the water fog area in the image, and the candidate frames w, h, x, y are obtained, where w and h are the relative width and height of the candidate frame (the relative width and height are respectively the width and height of the candidate frame divided by the width and height of the image), and x and y are the relative center coordinates of the candidate frame. The same image can contain multiple candidate frames. After the images in the collection, The image with the same water level as the annotated image is used as the real image after removing the water mist; annotate one by one The images in the set are found, and the corresponding paired images are found to complete the production of the data set; the data set data is randomly divided into a training set and a validation set in a ratio of 8:2.

[0020] S3. Construct a neural network training model for water mist removal. The model structure is as follows: Figure 2 As shown, it includes an encoder-decoder network and a discriminant network, where the encoder-decoder network structure is as follows Figure 3 The encoder skeleton network in the encoder-decoder network is as shown in Figure 4 shown. Figure 2-Figure 4 The ConvNeXt and downsampling modules described in Figure 5 shown. Figure 2-Figure 5 The meanings of the parameters are as follows: h, w, dim are the image feature height, width and number of channels respectively; k is the convolution kernel size; s is the convolution step size; p is the feature edge zero padding, where p=i means padding i columns of zeros around the feature, p=(i1,i2,i3,i4) means padding i1 rows of zeros above the feature, i2 rows of zeros below, i3 columns of zeros on the left, and i4 columns of zeros on the right; GELU is the GELU activation function, and its formula is ; The specific details of the model are as follows: a 960*960 RGB image with water fog is used as the input of the model, and an image without water fog is generated through the encoder-decoder network. The corresponding manually annotated candidate box area is cropped and scaled to 240*240 size. The generated image without water fog and the cropped and scaled image are used as input, and the discriminant network calculates and outputs the confidence of the generated image without water fog.

[0021] Specifically, the water mist removal neural network training model in step S3 is composed of input (960*960*3 RGB image), encoder-decoder network, and discriminant network; The encoder-decoder network is used to generate the water fog removed image from the original input image, which includes an encoder network and a decoder network. The encoder network consists of an encoder skeleton network and an atrous spatial pyramid pooling (ASPP) module.

[0022] The encoder skeleton network consists of an initialized convolution layer (consisting of a convolution layer with a kernel size k of 7, a stride s of 4, a zero padding method p of [2,1,2,1], and a layer normalization LN, where [2,1,2,1] means 2 layers of zero padding on the left and above the image, and 1 layer of zero padding on the bottom and right of the image), 3 ConvNeXt modules with 64 output channels (the ConvNeXt module includes a residual structure, a depthwise separable convolution layer with k of 7, s of 1, and p of 3, a layer normalization LN, a convolution layer with k of 1 and s of 1, a GELU activation function, a convolution layer with k of 1 and s of 1, a layer scaling, and a Drop Path layer), and its output feature is Drop The network consists of the feature stacking of the Path layer output features and the ConvNext module input features), a downsampling module (the downsampling module consists of layer normalization LN and a convolutional layer with k=7, s=2, and p=[3,2,3,2]), 9 ConvNeXt modules with an output channel of 128, a downsampling module, and 3 ConvNeXt modules with an output channel of 256. The encoder network outputs 240*240*64 low-dimensional features and 60*60*256 high-dimensional features in the 3rd ConvNeXt and 15th ConvNeXt modules, respectively.

[0023] The Atrous Spatial Pyramid Pooling (ASPP) module consists of input (high-dimensional features output by the encoder skeleton network), 5 branches, feature concatenation, 1*1 convolution (convolution with k=1, s=1, layer normalization LN, GELU activation function), and 4-fold bilinear interpolation. Branch 1 is a 1*1 convolution; branch 2 is a dilated convolution with k=3, s=1, p=6, and dilation factor d=6, layer normalization LN and GELU activation function; branch 3 is a dilated convolution with k=3, s=1, p=12, and d=12, layer normalization LN and GELU activation function; branch 4 is a dilated convolution with k=3, s=1, p=18, and d=18, layer normalization LN and GELU activation function; branch 5 is average pooling, 1*1 convolution, and 2-fold bilinear interpolation. The features output by the 5 branches are concatenated, and 1*1 convolution and 4-fold bilinear interpolation are input. The Atrous Spatial Pyramid Pooling (ASPP) module is used to collect image features of different scales, increase the network's receptive field, and enhance the network's ability to extract global image information.

[0024] The decoder network consists of input (low-dimensional features output by the encoder skeleton network), 1*1 convolution, feature concatenation (concatenating 1*1 convolution output features and features output by the encoder network), convolution with k=3, s=1, p=1, layer normalization LN and GELU activation functions, and 4-fold bilinear interpolation.

[0025] The discriminant network is used to assist the encoder-decoder network in training the ability to remove water mist. Its structure includes two branches. Branch 1 includes input (features output by the encoder-decoder), an initialized convolution layer (consisting of a convolution layer with a convolution kernel size k of 7, a step size s of 4, a zero-padding method p of [2,1,2,1], and a layer normalization LN), a ConvNeXt module with an output channel of 128, a downsampling module, a ConvNeXt module with an output channel of 256, and a downsampling module; Branch 2 includes input (an image cropped by the four parameters w, h, x, and y corresponding to the true values ​​of the features output by the encoder-decoder network), 1*1 convolution, a ConvNeXt module with an output channel of 128, a downsampling module, a ConvNeXt module with an output channel of 256, and a downsampling module. The output features of branch 1 and branch 2 are concatenated, and the input is a 1*1 convolution, a ConvNeXt module with an output channel of 512, a downsampling module, a ConvNeXt module with an output channel of 1024, a downsampling module, a global average pooling and a softmax layer.

[0026] S4. Training the water mist removal neural network training model: (1) Scale the training set data to 960*960 and normalize it, and crop and scale the manually labeled area in the training data set to 240*240 and normalize it; (2) Set the basic training parameter BATCH_SIZE to 8; (3) When the number of training rounds EPOCH is 200 or less, only train the encoder-decoder network, Adam is the training optimizer, the initial learning rate is 0.01, and its loss function is set to mean square error loss , whose formula is 2 .in Generate the pixel value of the i-th pixel for the input image annotation area through neural network prediction, is the true pixel value of the i-th pixel, and the parameter P1 with the best performance in the validation set at this stage is stored; (4) When the training round number EPOCH∈(200,500], the training model is trained as a whole, the encoder-decoder network loads the parameter P1, Adam is used as the training optimizer, the initial learning rate is 0.001, and its loss function is set to + ,in , where D is the discriminator, i.e., the discriminant network, G is the generator, i.e., the encoder-decoder network, x is the real image, G(z) is the image generated by the encoder-decoder network, and the parameter P2 with the best performance in the validation set at this stage is stored; S5. Use the trained water fog removal model to remove water fog from the image: (1) Obtain the water level image captured by the water collection well monitoring camera; (2) Scale the image to 960*960 size; (3) Normalize the scaled image; (4) Use the encoder-decoder network to load its corresponding parameters in P2; (4) Input the normalized image into the encoder-decoder network, which outputs the defogged image; (5) Scale the defogged image to its original size to complete the defogging process.

[0027] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A method for removing water mist from a video monitoring image of a water collection well in a hydropower station, characterized in that: It includes the following steps: S1. Capture the water collection well image through the surveillance camera, including the image set with water mist and a collection of images without water mist ; S2. Image collection and Perform manual annotation to create data sets; S3, building a neural network training model for water mist removal; S4, training a neural network training model for water mist removal; S5. Use the trained water mist removal model to remove water mist from the image.

2. A method for removing water mist from a water collection well video monitoring image of a hydropower station according to claim 1, characterized in that: The specific process of step S2 is as follows: Use labelimg to label Annotate the images in the collection, generate candidate frames for the water fog area in the image, and obtain the candidate frames w, h, x, y, where w and h are the relative width and height of the candidate frame, the relative width and height are the width and height of the candidate frame divided by the width and height of the image, and x and y are the relative center coordinates of the candidate frame. The same image contains multiple candidate frames. After marking a After the images in the collection, The image with the same water level as the annotated image is used as the real image after removing the water mist; Mark one by one The images in the collection are searched and the corresponding paired images are found to complete the dataset creation; The data set is randomly divided into a training set and a validation set in a ratio of 8:

2.

3. The method for removing water mist from a water collection well video monitoring image of a hydropower station according to claim 1, characterized in that: The water mist removal neural network training model in step S3 is composed of input, encoder-decoder network, and discriminant network; The encoder-decoder network is used to generate the image after water fog removal from the original input image, which includes an encoder network and a decoder network; the encoder network includes an encoder skeleton network and a void space pyramid pooling module.

4. A method for removing water mist from a water collection well video monitoring image of a hydropower station according to claim 3, characterized in that: The encoder skeleton network consists of: Initialize the convolutional layer, which consists of a convolutional layer with a kernel size k of 7, a stride s of 4, a zero padding method p of [2,1,2,1], and a layer normalization LN. [2,1,2,1] means 2 layers of zero padding on the left and above the image, and 1 layer of zero padding on the bottom and right of the image. 3 ConvNeXt modules with 64 output channels, where the ConvNeXt module contains a residual structure; a depth-wise separable convolutional layer with k=7, s=1, and p=3; layer normalization LN; a convolutional layer with k=1 and s=1; a GELU activation function; a convolutional layer with k=1 and s=1; layer scaling and a Drop Path layer, whose output features are the feature stacking of the Drop Path layer output features and the ConvNext module input features; Downsampling module,The downsampling module consists of layer normalization LN and convolutional layers with k=7, s=2, and p=[3,2,3,2]; 9 ConvNeXt modules with 128 output channels, a downsampling module, and 3 ConvNeXt modules with 256 output channels. The encoder network outputs 240*240*64 low-dimensional features and 60*60*256 high-dimensional features in the 3rd ConvNeXt and 15th ConvNeXt modules respectively; The atrous spatial pyramid pooling module includes: Input, i.e., the high-dimensional features output by the encoder skeleton network; 5 branches; feature concatenation and 1*1 convolution, which includes convolution with k=1 and s=1, layer normalization LN, GELU activation function; 4x bilinear interpolation; Among them, branch 1 is 1*1 convolution; branch 2 is a dilated convolution with k=3, s=1, p=6, and expansion coefficient d=6, and layer-normalized LN and GELU activation functions; branch 3 is a dilated convolution with k=3, s=1, p=12, and d=12, and layer-normalized LN and GELU activation functions; branch 4 is a dilated convolution with k=3, s=1, p=18, and d=18, and layer-normalized LN and GELU activation functions; branch 5 is average pooling, 1*1 convolution, and 2x bilinear interpolation; the features output by the 5 branches are spliced, input into 1*1 convolution and 4x bilinear interpolation; the dilated spatial pyramid pooling module is used to collect image features of different scales, increase the network's receptive field, and enhance the network's ability to extract global image information; The decoder network includes: input, that is, the low-dimensional features output by the encoder skeleton network; 1*1 convolution; feature splicing, that is, splicing the 1*1 convolution output features and the features output by the encoder network; convolution and layer normalization LN and GELU activation functions with k=3, s=1, and p=1; 4x bilinear interpolation.

5. The method for removing water mist from a water collection well video monitoring image of a hydropower station according to claim 3, characterized in that: The discriminant network is used to assist the encoder-decoder network in training the ability to remove water mist. Its structure includes two branches. Branch 1 includes input, that is, the features of the encoder-decoder output; the convolution layer is initialized, which consists of a convolution layer with a convolution kernel size k of 7, a step size s of 4, a zero padding method p of [2,1,2,1] and a layer normalization LN; a ConvNeXt module with an output channel of 128 and a downsampling module; a ConvNeXt module with an output channel of 256 and a downsampling module; Branch 2 includes input, that is, the encoder-decoder network output The image is cropped by the four parameters w, h, x, and y of the true values ​​corresponding to the features output; 1*1 convolution; ConvNeXt module with an output channel of 128, downsampling module; ConvNeXt module and downsampling module with an output channel of 256; the output features of branch 1 and branch 2 are concatenated, and 1*1 convolution, ConvNeXt module with an output channel of 512, downsampling module, ConvNeXt module with an output channel of 1024, downsampling module, global average pooling and softmax layer are input.

6. The method for removing water mist from a water collection well video monitoring image of a hydropower station according to claim 3, characterized in that: When constructing the water mist removal neural network training model in step S3, a water mist RGB image with a size of 960*960 is used as the input of the model, and an image without water mist is generated through the encoder-decoder network. The corresponding manually marked candidate frame area is cropped and scaled to a size of 240*240. The generated image without water mist and the cropped and scaled image are used as input, and the discriminant network calculates and outputs the confidence of the generated image without water mist.

7. The method for removing water mist from a water collection well video monitoring image of a hydropower station according to claim 1, characterized in that: The specific process of step S4 is as follows: 4.

1. Scale the training set data to 960*960 and normalize it. Crop and scale the manually annotated area in the training data set to 240*240 and normalize it. 4.

2. Set the basic training parameter BATCH_SIZE to 8; 4.

3. When the number of training rounds EPOCH is 200 or less, only the encoder-decoder network is trained, Adam is the training optimizer, the initial learning rate is 0.01, and its loss function is set to mean square error loss , whose formula is 2 ;in Generate the pixel value of the i-th pixel for the input image annotation area through neural network prediction, is the true pixel value of the i-th pixel, and stores the parameter P1 that performs best in the validation set at this stage; 4.

4. When the training round number EPOCH∈(200,500], the training model is trained as a whole, the encoder-decoder network loads the parameter P1, Adam is used as the training optimizer, the initial learning rate is 0.001, and its loss function is set to + ,in , where D is the discriminator, i.e., the discriminant network, G is the generator, i.e., the encoder-decoder network, x is the real image, G(z) is the image generated by the encoder-decoder network, and the parameter P2 with the best performance on the validation set at this stage is stored.

8. The method for removing water mist from a water collection well video monitoring image of a hydropower station according to claim 1, characterized in that: The specific process of step S5 is as follows: 5.

1. Obtain the water level image captured by the water collection well monitoring camera; 5.

2. Scale the image to 960*960 size; 5.

3. Normalize the scaled image; 5.

4. Use the encoder-decoder network to load the corresponding parameters in P2; 5.

5. Input the normalized image into the encoder-decoder network, which outputs the dehazed image; 5.

6. Scale the dehazed image to its original size to complete the dehazing process.

Images