A method for detecting water quality transparency using a Secchi disk without a water gauge based on deep learning

Abstract

The present invention discloses a method for detecting the water quality transparency of a Secchi disk without a water gauge based on deep learning, comprising the following steps: Step 1, obtaining a video of measuring the water quality transparency of the Secchi disk; Step 2, determining the initial position of the Secchi disk; then using the SiamFC++ network to track the position of the Secchi disk in other frames of the video according to the initial position of the Secchi disk, and segmenting the tracking result of each frame as the segmentation result; Step 3, using a discriminator to perform data cleaning on the segmentation result obtained in Step 2 to obtain the critical position of the Secchi disk, and then removing the part after the critical position of the Secchi disk from the segmentation result; Step 4, extracting features from the remaining part of the segmentation result obtained in Step 3, and then using the recurrent network Bi-Lstm to calculate the water quality transparency calculation result based on the extracted features. The present invention does not require observing the water gauge, avoiding many inconveniences caused by reading the water gauge, and greatly improving the measurement speed and accuracy.

Description

Technical Field

[0001] The present invention relates to the field of water quality transparency detection methods, and specifically to a waterless scale Secchi disk water quality transparency detection method based on deep learning. Background Art

[0002] Water quality transparency is a commonly used indicator to measure the quality of water, and the Secchi disk method has always been a commonly used method for water quality transparency detection. The tester can obtain the water quality transparency value by using the human eye to distinguish the moment when the Secchi disk "disappears" in the water and reading the water scale reading at the same time. Although the traditional detection method using the Secchi disk method is simple and practical, it is greatly affected by the tester's eyesight and subjective judgment during the detection process, with relatively large uncertainty. In addition, considering many external factors, the transparency value obtained by detection will deviate.

[0003] In recent years, image processing technology has been continuously improved, and more efficient deep learning methods have emerged in an endless stream, enabling the computer to quickly identify various objects in the image and classify them, which not only realizes automatic detection but also avoids various subjective factors that may be encountered in human eye recognition. The development of such technologies provides strong support for the detection efficiency and the reliability of the results.

[0004] In the existing methods for measuring transparency based on machine learning, such as the Chinese patent document with publication number CN113252614A discloses an intelligent Secchi disk and water scale recognition technology based on image processing. This method needs to rely on the reading of the water scale to measure the water quality transparency. Therefore, when observing the Secchi disk, it is also necessary to obtain the water scale image, and then calculate and obtain the water scale reading based on the water scale image, which brings inconvenience to the measurement, and errors are also easily introduced during the calculation of the water scale reading, resulting in deviation of the final calculation result of the water quality transparency. At the same time, the determination process of the critical position of the Secchi disk in this method is relatively complex, so it will lead to a long time-consuming calculation of the water quality transparency and ultimately reduce the detection efficiency of the water quality transparency. Summary of the Invention

[0005] The purpose of the present invention is to provide a waterless scale Secchi disk water quality transparency detection method based on deep learning to solve the problem of the need to calculate the water scale reading when measuring water quality transparency based on the Secchi disk and water scale in the prior art.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A waterless scale Secchi disk water quality transparency detection method based on deep learning includes the following steps:

[0008] Step 1, obtain a video of measuring the water quality transparency of the Secchi disk;

[0009] Step 2: Perform image processing and recognition on the video obtained in Step 1 to obtain the position of the Secchi disk in the first frame of the video as the initial position of the Secchi disk; then use the SiamFC++ network based on 2D convolution to track the position of the Secchi disk in other frames of the video according to the initial position of the Secchi disk to obtain the tracking result, and segment the tracking result of each frame from the corresponding frame of the video as the segmentation result;

[0010] Step 3: Use a discriminator based on 3D convolution to perform data cleaning on the segmentation result obtained in Step 2 to obtain the critical position of the Secchi disk, and then remove the part after the critical position of the Secchi disk from the segmentation result, and keep the remaining part in the segmentation result;

[0011] Step 4: Extract features from the remaining part of the segmentation result obtained in Step 3, and then use the recurrent network Bi-Lstm to calculate the water quality transparency calculation result based on the extracted features.

[0012] Further, the process of Step 2 is as follows;

[0013] Step 2-1: Use the selectROI function in OpenCV to perform image processing and recognition on the video, and initialize the coordinates of the detection box of the Secchi disk in the first frame of the video through the selectROI function in OpenCV as the initial position of the Secchi disk;

[0014] Step 2-2: Input other frames of the video into the SiamFC++ network frame by frame, and the SiamFC++ network tracks the area framed by the detection box of the Secchi disk in other frames of the video according to the initial position of the Secchi disk as the tracking result; then based on the coordinates of the detection box of the Secchi disk in other frames of the video, segment the tracking result from other frames of the video respectively as the segmentation result.

[0015] Further, in Step 2-2, use the area framed by the detection box of the Secchi disk with the highest confidence in other frames of the video obtained by tracking as the tracking result, and segment the tracking result as the segmentation result based on the coordinates of the detection box of the Secchi disk with the highest confidence.

[0016] Further, in Step 2, the SiamFC++ network is built with ShuffleNetV2 as the backbone network; for each input image, the SiamFC++ outputs a coordinate matrix with a row size of H and a column size of W and a confidence matrix, so there are H*W coordinates of detection boxes and H*W confidences of detection boxes; the coordinates of each detection box are recorded as [x i , y i , h i , w i , where x i , y irespectively represent the coordinates of the upper left corner of the detection box, h i and w i respectively represent the height and width of the detection box, and 1 ≤ i ≤ H * W; the confidence of each detection box is between 0 and 1.

[0017] Furthermore, in the SiamFC++ network, both the row size H and the column size W are set to 17.

[0018] Furthermore, the process of step 3 is as follows:

[0019] Step 3-1: Assume that there are N segmentation results obtained in step 2, and every consecutive K of them are grouped to obtain multiple groups. These multiple groups are used as the input of the discriminator, and the discriminator outputs N - K + 1 probability values, and the range of each probability value is (0, 1);

[0020] Step 3-2: Determine the critical position of the Secchi disk according to the N - K + 1 probability values obtained in step 3-1, where:

[0021] If there is a probability value of 0.5 among the N - K + 1 probability values, then the position where 0.5 appears for the first time is used as the critical position of the Secchi disk;

[0022] If there is no probability value of 0.5 among the N - K + 1 probability values, then the N - K + 1 probability values are stored in a one-dimensional array, and a sliding window slides from the starting position of the one-dimensional array with a fixed step size. Each time it slides, the number of probability values greater than 0.5 and less than 0.5 in the sliding window is counted. When the number of probability values greater than 0.5 is greater than the number of probability values less than 0.5, the center position of the sliding window at this time is the critical position of the Secchi disk;

[0023] Step 3-3: Remove the part after the critical position of the Secchi disk from the segmentation result, and keep the remaining part in the segmentation result.

[0024] Furthermore, the discriminator includes 5 layers of networks, where:

[0025] The first layer of network includes 64 3D convolution kernels of size [5, 5, 1], the convolution stride is [1, 1, 1], the input size of the first layer of network is 56 * 56 * 3 * 8, and the output size is 56 * 56 * 64 * 8;

[0026] The second layer of network includes 128 3D convolution kernels of size [3, 3, 1], the convolution stride is [2, 2, 1], the input size of the second layer of network is 56 * 56 * 64 * 8, and the output size is 28 * 28 * 128 * 8;

[0027] The third - layer network includes 256 3D convolutional kernels of size [3, 3, 1], with a convolutional stride of [2, 2, 1]. The input size of the third - layer network is 28*28*128*8, and the output size is 14*14*256*8;

[0028] The fourth - layer network includes 512 3D convolutional kernels of size [3, 3, 1], with a convolutional stride of [2, 2, 1]. The input size of the fourth - layer network is 14*14*256*8, and the output size is 7*7*512*8;

[0029] The fifth - layer network includes an average pooling layer, a fully - connected layer, and an activation layer. The activation function of the activation layer uses the sigmoid function. The input size of the fifth - layer network is 7*7*512*8, and the output size is 1*1*1*1.

[0030] Further, in step 4, first, the remaining part in the segmentation result of step 3 is converted into a grayscale image, then the interval where each pixel in the grayscale image is located is statistically analyzed, and a one - dimensional feature vector is established based on the statistical result; then the one - dimensional feature vector is normalized, and the normalization result is used as the extracted feature.

[0031] Further, in step 4, the recurrent network Bi - Lstm takes the extracted feature as input. The output of the recurrent network Bi - Lstm is connected to a two - layer fully - connected network, and finally, the water quality transparency calculation result is output by the fully - connected network.

[0032] The present invention proposes a transparency measurement algorithm based on a convolutional neural network and a recurrent neural network. By processing the Secchi disk video through the convolutional neural network and the recurrent neural network, it can quickly determine the critical position of the Secchi disk and directly calculate the value of water quality transparency. It does not require the cooperation of a water gauge. Therefore, it can avoid observing and calculating the water gauge reading. Only the Secchi disk video is needed to measure the water quality transparency calculation result, which has the advantages of low cost, fast measurement, small error, and objective water quality transparency calculation result, and has high application value. Description of the Drawings

[0033] Figure 1 It is a schematic flowchart of a method for detecting water quality transparency of a Secchi disk without a water gauge based on deep learning in an embodiment of the present invention.

[0034] Figure 2 It is a Secchi disk segmentation result picture and its corresponding grayscale histogram in an embodiment of the present invention, where (a) is the Secchi disk segmentation result picture and (b) is the grayscale histogram.

[0035] Figure 3 It is a partial tracking effect picture during the process of SiamFC++ tracking the Secchi disk in an embodiment of the present invention.

[0036] Figure 4 This is a partial picture of the Secchi disk segmentation result in the embodiment of the present invention.

[0037] Figure 5 This is a partial picture of the Secchi disk segmentation result after data cleaning in the embodiment of the present invention.

[0038] Figure 6 This is a schematic diagram of calculating water quality transparency by Bi-Lstm in the embodiment of the present invention. Detailed implementation manners

[0039] To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described below in conjunction with embodiments and their accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the described embodiments without creative efforts fall within the scope of protection of the present invention.

[0040] Unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meaning understood by those of ordinary skill in the art to which the present invention belongs. The terms "including" or "comprising" and the like used in the present invention mean that the elements or objects appearing before this word cover the elements or objects listed after this word and their equivalents, without excluding other elements or objects. The terms "connected" or "coupled" and the like are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms "upper", "lower", "left", "right" and the like are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0041] As Figure 1 shown, the water quality transparency detection method of the Secchi disk without water gauge based on deep learning in this embodiment includes the following steps:

[0042] S100. Obtain the Secchi disk water quality transparency measurement video.

[0043] In this embodiment, a camera is used to shoot the Secchi disk to obtain a Secchi disk lifting video with a relatively stable and clear picture as the Secchi disk water quality transparency measurement video. Then, the video shot by the camera is transmitted to a computer through the cloud or directly for processing, and the computer completes the subsequent steps.

[0044] S200. Secchi disk tracking and segmentation:

[0045] Use the SiamFC++ network based on 2D convolution to track the Secchi disk in the video obtained in step S100, and segment the tracking result as the segmentation result. The specific process is as follows:

[0046] (S2001) Use the selectROI function in OpenCV to perform image detection and recognition on the video obtained in step S100. Initialize the detection box coordinates of the Secchi disk in the first frame of the video through the selectROI function in OpenCV as the initial position of the Secchi disk.

[0047] (S2002) Then, input other frames of the video into the SiamFC++ network frame by frame. The SiamFC++ network tracks the region framed by the Secchi disk detection box with the highest confidence in other frames of the video based on the initial position of the Secchi disk as the tracking result, and separates the tracking result from other frames of the video according to the coordinates of the detection box with the highest confidence as the segmentation result.

[0048] In step S200 of this embodiment, the SiamFC++ network based on 2D convolution is used as the tracker. The SiamFC++ network is built with ShuffleNetV2 as the backbone network, and the Secchi disk in the video is tracked through the SiamFC++ network. Some effect pictures of the SiamFC++ tracking the Secchi disk in this embodiment are as Figure 3 shown.

[0049] The SiamFC++ network can automatically detect the position of the Secchi disk in other frames according to the initial position of the Secchi disk in the video. For each input picture, the SiamFC++ outputs a coordinate matrix and a confidence matrix with a row size of H and a column size of W. Therefore, there are H*W coordinates of detection boxes and H*W confidences of detection boxes; the coordinates of each detection box are recorded as [x i , y i , h i , w i , where x i , y i respectively represent the coordinates of the upper left corner of the detection box, h i , w i respectively represent the height and width of the detection box, and 1 ≤ i ≤ H*W; the confidence of each detection box is between 0 and 1.

[0050] When the coordinates of the detection box with the highest confidence in each frame of the picture are found according to the output of the SiamFC++ network, the detection box area can be separated from the corresponding frame of the video according to the coordinates of the detection box with the highest confidence as the segmentation result. When the Secchi disk in the picture is in a visible state, the segmentation result contains the Secchi disk; when the Secchi disk in the picture is in an invisible state, there is no Secchi disk in the segmentation result. Some Secchi disk segmentation result pictures in this embodiment are as Figure 4 shown.

[0051] S300. Cleaning the segmentation result data:

[0052] Use a discriminator based on 3D convolution to clean the segmentation result obtained in step S200, and remove the part after the critical position of the Secchi disk from the segmentation result, and keep the remaining part of the segmentation result. The specific steps are as follows:

[0053] (S3001). Adopt a discriminator based on 3D convolution; this discriminator has 5 layers of networks, and the network parameters of each layer of the discriminator are shown in Table 1, and Table 1 is as follows:

[0054] Table Network parameters of each layer of the discriminator

[0055] Name of the layer Input size Output size Parameter Stage1 56*56*3*8 56*56*64*8 64 3D convolutional kernels of size [5, 5, 1], convolution stride [1, 1, 1] Stage2 56*56*64*8 28*28*128*8 128 3D convolutional kernels of size [3, 3, 1], convolution stride [2, 2, 1] Stage3 28*28*128*8 14*14*256*8 256 3D convolutional kernels of size [3, 3, 1], convolution stride [2, 2, 1] Stage4 14*14*256*8 7*7*512*8 512 3D convolutional kernels of size [3, 3, 1], convolution stride [2, 2, 1] Stage5 7*7*512*8 1*1*1*1 Average pooling layer, fully connected layer, activation layer (using sigmoid function as the activation function)

[0056] The input of the discriminator is K Secchi disk segmentation result pictures (in this example, K = 8), and the output is a probability value between (0, 1). This probability value reflects the state of the Secchi disk. When the probability value is close to 0, it means the Secchi disk is in a visible state; when the probability value is close to 1, it means the Secchi disk is in an invisible state; when the probability value is exactly equal to 0.5, it indicates that the Secchi disk is in a critical state between visible and invisible.

[0057] (S3002). Determine the critical position of the Secchi disk:

[0058] Suppose there are N Secchi disk segmentation result pictures in total in step S200. Take every consecutive K pictures as a group (in this embodiment, K = 8) as the input and send it into the discriminator, and then the discriminator outputs a probability value. A total of N - K + 1 probability values can be obtained.

[0059] If among the N - K + 1 probability values, there is a probability value of 0.5, then take the position where 0.5 appears for the first time as the critical position of the Secchi disk.

[0060] If among the N - K + 1 probability values, there is no probability value of 0.5, then determine the critical position through the following method: Store the N - K + 1 probability values in a one-dimensional array, use a sliding window with a length of 5, start sliding from the starting position of the array with a fixed step size of 1, and each time it slides, count the number of probability values greater than 0.5 and less than 0.5 in the sliding window. When the number of probability values greater than 0.5 is greater than the number of probability values less than 0.5, the center position of the sliding window at this time is the critical position of the Secchi disk.

[0061] (S3003). After determining the critical position of the Secchi disk, remove and discard the part after the critical position of the Secchi disk from the segmentation result obtained in step S200, and keep the remaining part of the segmentation result.

[0062] In this embodiment, after the tracking and segmentation in step S200, a series of Secchi disk segmentation result pictures will be obtained. In these pictures, some Secchi disks are in a visible state, some are in an invisible state, and some are in a critical state between visible and invisible. The pictures with invisible Secchi disks are useless for calculating water quality transparency. Therefore, it is necessary to determine in which picture the critical position of the Secchi disk is. There are no Secchi disks in the pictures after the critical position, which are data that need to be cleaned, and the pictures before the critical position are data that need to be retained. Some of the Secchi disk segmentation result pictures after data cleaning in this embodiment are as Figure 5 shown.

[0063] S400. After the data cleaning in step S300, there are M Secchi disk segmentation result pictures left. Feature extraction is performed on the remaining part of the segmentation results, and the water quality transparency calculation result is calculated using the recurrent network Bi-Lstm.

[0064] In the video, the pixels on the Secchi disk and the depth at the position where the Secchi disk is located change with time. Therefore, the temporal relationship between the Secchi disk pixels and the Secchi disk depth can be established through a recurrent neural network. When the Secchi disk is at the critical position, the Secchi disk depth at this time is the transparency. The recurrent neural network uses the recurrent network Bi-Lstm.

[0065] The process of calculating water quality transparency based on Bi-Lstm in this step S400 is as Figure 6 shown, and the specific steps are as follows:

[0066] (S4001) Feature extraction:

[0067] As Figure 2 shown, the remaining part of the Secchi disk segmentation result pictures is converted from an RGB picture to a grayscale picture, and the histogram of the grayscale picture is made. Each pixel point of the grayscale picture is in the interval [0, 255]. The interval [0, 255] is divided into 32 equal-length sub-intervals. The number of pixels falling into each sub-interval is counted, and a one-dimensional feature vector with a length of 32 is established. The number of pixels counted in each sub-interval is sequentially stored in this one-dimensional feature vector.

[0068] Then, the one-dimensional feature vector is normalized using L2 normalization. The calculation formula of L2 normalization is shown in formula (1):

[0069] (1)

[0070] In formula (1), Y represents the normalized feature vector, which is the feature to be extracted in this step; X represents the feature vector before normalization, which is the one-dimensional feature vector.

[0071] (S4002)Calculation of water quality transparency:

[0072] According to step (S4001), feature extraction is performed on all Secchi disk segmentation results after data cleaning. The extracted features are sent into the recurrent network Bi-Lstm, and then the results output by the recurrent network Bi-Lstm are sent into a two-layer fully connected network. Finally, the fully connected network outputs the calculation result of water quality transparency.

[0073] S500. Output the calculation result of water quality transparency obtained in step S400.

[0074] The embodiments described in the present invention are only descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various variations and improvements made by those skilled in the art to the technical solutions of the present invention should fall within the protection scope of the present invention. The technical content claimed by the present invention has been fully recorded in the claims.

Claims

1. A method for detecting water quality transparency using a Secchi disk without a water gauge based on deep learning, characterized in that, It includes the following steps: Step 1: Obtain the video of Secchi disk water quality transparency measurement; Step 2: Perform image processing and recognition on the video obtained in Step 1 to obtain the position of the Secchi disk in the first frame of the video as the initial position of the Secchi disk; then use the SiamFC++ network based on 2D convolution to track the position of the Secchi disk in other frames of the video according to the initial position of the Secchi disk to obtain the tracking result, and segment the tracking result of each frame from the corresponding frame of the video as the segmentation result; Step 3: Use a discriminator based on 3D convolution to perform data cleaning on the segmentation result obtained in Step 2 to obtain the critical position of the Secchi disk, and then remove the part after the critical position of the Secchi disk from the segmentation result, and keep the remaining part in the segmentation result; Step 4: Extract features from the remaining part in the segmentation result obtained in Step 3, and then use the recurrent network Bi-Lstm to calculate the water quality transparency calculation result based on the extracted features; The process of Step 3 is as follows: Step 3-1: Assume that there are N segmentation results obtained in Step 2, and take every consecutive K of them as a group to obtain multiple groups, and use the obtained multiple groups as the input of the discriminator. The discriminator outputs N-K+1 probability values, and the range of each probability value is (0,1); Step 3-2: Determine the critical position of the Secchi disk according to the N-K+1 probability values obtained in Step 3-1, where: If there is a probability value of 0.5 among the N-K+1 probability values, take the position where 0.5 appears for the first time as the critical position of the Secchi disk; If there is no probability value of 0.5 among the N-K+1 probability values, store the N-K+1 probability values in a one-dimensional array, and use a sliding window to slide from the starting position of the one-dimensional array with a fixed step length. Each time it slides, count the number of probability values greater than 0.5 and less than 0.5 in the sliding window. When the number of probability values greater than 0.5 is greater than the number of probability values less than 0.5, the center position of the sliding window at this time is the critical position of the Secchi disk; Step 3-3: Remove the part after the critical position of the Secchi disk from the segmentation result, and keep the remaining part in the segmentation result.

2. The method for detecting water quality transparency using a Secchi disk without a water gauge based on deep learning according to claim 1, characterized in that, The process of Step 2 is as follows; Step 2-1: Use the selectROI function in OpenCV to perform image processing and recognition on the video, and initialize the detection box coordinates of the Secchi disk in the first frame of the video through the selectROI function in OpenCV as the initial position of the Secchi disk; Step 2-2: Input each frame of the other frames of the video into the SiamFC++ network one by one. The SiamFC++ network tracks the area framed by the detection box of the Secchi disk in other frames of the video according to the initial position of the Secchi disk as the tracking result; then based on the detection box coordinates of the Secchi disk in other frames of the video, segment the tracking result from other frames of the video respectively as the segmentation result.

3. The method for detecting water quality transparency using a Secchi disk without a water gauge based on deep learning according to claim 2, characterized in that, In Step 2-2, take the area framed by the detection box of the Secchi disk with the highest confidence in other frames of the video obtained by tracking as the tracking result, and segment the tracking result as the segmentation result based on the coordinates of the detection box of the Secchi disk with the highest confidence.

4. The method for detecting water quality transparency using a Secchi disk without a water gauge based on deep learning according to claim 1 or 2 or 3, characterized in that, In step 2, the SiamFC++ network is built with ShuffleNetV2 as the backbone network; for each input image, the SiamFC++ outputs a coordinate matrix with a row size of H and a column size of W and a confidence matrix, so there are H*W coordinates of detection boxes and H*W confidences of detection boxes; the coordinates of each detection box are denoted as [x i , y i , h i , w i , where x i , y i represent the coordinates of the upper left corner of the detection box respectively, h i , w i represent the height and width of the detection box respectively, and 1 ≤ i ≤ H*W; the confidence of each detection box is between 0 and 1.

5. The method for detecting water quality transparency using a Secchi disk without a water gauge based on deep learning according to claim 4, characterized in that, In the described SiamFC++ network, both the row size H and the column size W are set to 17.

6. The method for detecting water quality transparency using a Secchi disk without a water gauge based on deep learning according to claim 1, characterized in that, The discriminator includes a 5-layer network, where: The first layer network includes 64 3D convolution kernels of size [5, 5, 1], with a convolution stride of [1, 1, 1]. The input size of the first layer network is 56 * 56 * 3 * 8, and the output size is 56 * 56 * 64 * 8; The second layer network includes 128 3D convolution kernels of size [3, 3, 1], with a convolution stride of [2, 2, 1]. The input size of the second layer network is 56 * 56 * 64 * 8, and the output size is 28 * 28 * 128 * 8; The third layer network includes 256 3D convolution kernels of size [3, 3, 1], with a convolution stride of [2, 2, 1]. The input size of the third layer network is 28 * 28 * 128 * 8, and the output size is 14 * 14 * 256 * 8; The fourth layer network includes 512 3D convolution kernels of size [3, 3, 1], with a convolution stride of [2, 2, 1]. The input size of the fourth layer network is 14 * 14 * 256 * 8, and the output size is 7 * 7 * 512 * 8; The fifth layer network includes an average pooling layer, a fully connected layer, and an activation layer. The activation function of the activation layer uses the sigmoid function. The input size of the fifth layer network is 7 * 7 * 512 * 8, and the output size is 1 * 1 * 1 * 1.

7. The method for detecting water quality transparency using a Secchi disk without a water gauge based on deep learning according to claim 1, characterized in that, In step 4, first, the remaining part in the segmentation result of step 3 is converted into a grayscale image, and then the interval where each pixel in the grayscale image is located is statistically analyzed, and a one-dimensional feature vector is established based on the statistical result; then The one-dimensional feature vector is normalized, and the normalization result is used as the extracted feature.

8. The method for detecting water quality transparency using a Secchi disk without a water gauge based on deep learning according to claim 1 or 7, characterized in that, In step 4, the recurrent network Bi-Lstm uses the extracted feature as the input. The output of the recurrent network Bi-Lstm is connected to a two-layer fully connected network, and finally, the water quality transparency calculation result is output by the fully connected network.

Images