CN-122023883-A - Intelligent recognition method for ice water characteristics and ice sealing rate under low-contrast scene

Abstract

The invention relates to an intelligent recognition method for ice water characteristics and ice sealing rate under a low-contrast scene. The invention adopts a deformable convolution layer and an attention mechanism to guide the sampling points to extend towards the outline of the ice body, expands the sampling range to lead the sampling points to extend towards the outline of the ice body, so that the convolution kernel can break through the limitation of fixed shapes, further improves the recognition capability in the aspects of the outline and detailed characteristics of flowing ice, and further captures the important characteristics of ice and water in the two layers of ice and water color channels and spatial distribution, further captures the important characteristics of ice and water, and further improves the recognition precision of the ice sealing rate in a low-contrast scene.

Inventors

• FU HUI
• LI JIAZHEN
• GUO YONGXIN
• Chen Yuzhuang
• SU TAO
• GUO XINLEI
• SHEN YANQING
• XU HAIQING
• WANG TAO
• WANG JUN
• ZHANG YI
• LI ZHONGLIN
• PAN JIAJIA

Assignees

• 中国水利水电科学研究院

Dates

Publication Date

20260512

Application Date

20260109

Claims (1)

1. 1. The intelligent recognition method for the ice water characteristics and the ice sealing rate under the low-contrast scene is characterized by comprising the following steps of: The method comprises the steps of 1, acquiring video data of a monitoring camera, setting a time interval threshold t, intercepting an image frame from a read video at the threshold t, and then performing image format conversion and normalization size preprocessing; The method comprises the steps of marking data, finely dividing ice body and water body areas, generating a single-channel gray scale reference image, dividing the data, and classifying images in the data into a training set, a verification set and a test set according to the proportion of 75%, 20% and 5%; step 3, model construction and training, which comprises model construction and model training: (1) Model construction: The method comprises the steps of (1.1) extracting image features, namely extracting ice and water features in parallel by using a deep and shallow double-branch extraction network before a downsampling stage based on a U-Net model, sequentially inputting images into a conventional convolution layer with 3 x 3 convolution kernel sizes to realize feature information extraction, focusing on the whole outline of ice and water in the images by using a deep branch network, embedding an attention mechanism module behind each convolution layer by using 2 conventional convolution layers with 3 x 3 convolution kernel sizes, wherein the attention mechanism module comprises a channel attention mechanism for excavating important features of the ice and the water in a color channel dimension and a spatial attention mechanism for identifying positions of the important features in the images; The operation of the channel attention mechanism comprises the following substeps: step 311, obtaining the average value and the maximum value of all pixels of each channel through global average pooling and maximum pooling calculation; A sub-step 312 of generating a weight value between 0 and 1 for each channel through the Sigmoid function, wherein the closer the weight value is to 1, the more important the channel is for distinguishing ice and water interfaces; a sub-step 313 of amplifying the important channel features and suppressing the secondary channel features by multiplying the weight values with the pixel values of the original feature map; The operation of the spatial attention mechanism comprises the following substeps: step 321, calculating the average value and the maximum value of all channels in each spatial position in the feature map to obtain two single-channel maps which are respectively converged with different information so as to obtain which spatial positions show high response on all the feature channels; a substep 322, splicing the two images to generate a single spatial importance map; A substep 323 of normalizing each value in the spatial importance map to 0-1 by a Sigmoid function, wherein the position with higher value is more critical; A substep 324 of multiplying the map with the feature map after the channel attention treatment, thereby enhancing the feature response of the ice-water boundary region and weakening the irrelevant background region; The method comprises the following steps of (1.2) downsampling, wherein feature information after a double-branch extraction network can realize integration of feature information, dimension unification and feature diagram size clipping through a feature fusion module feature fusion module, wherein a downsampling stage comprises 5 deformable convolution modules, each deformable convolution module comprises 1 deformable convolution layer with a convolution kernel size of 3 multiplied by 3 and 2 deformable convolution layers with a convolution kernel size of 1 multiplied by 1, and an attention mechanism module is embedded behind each deformable convolution module; (1.3) upsampling the upsampling stage comprising 4 conventional convolution modules, each comprising 1 convolution layer having a convolution kernel size of 2 x 2, a convolution kernel size of 2 x 3, and a convolution kernel size of 1 x 1, respectively; the flow of the conventional convolution module is as follows: step 331, the feature map is passed through a convolution layer with a convolution kernel size of 2×2 to restore the resolution of the feature map to the original resolution; step 332, inputting the feature map into a convolution layer with a convolution kernel size of 3×3, and performing feature channel dimension reduction operation through a convolution layer with a convolution kernel size of 1×1, so as to simplify the calculation amount; (2) Model training: the training process calculates the average value of the cross-over ratio after the identification of the verification set image in real time, and triggers the early stop system to stop training when the cross-over ratio is continuously repeated for a plurality of times and is not lifted; And 4, model calling and ice sealing rate identification, which comprises the following sub-steps: a sub-step 401 of loading the trained optimal model parameters to switch the model to an inference mode and deploying the model into the adaptive computing equipment; A sub-step 402 of performing preprocessing operation completely consistent with the training set on the newly acquired ice image, including scaling the image to 1280×720 pixel resolution, converting to RGB three-channel format, normalizing the pixel values to 0-1 interval, and adjusting to tensor dimension of [1,3,720,1280 ]; A sub-step 403 of extracting image features through a dual-branch feature extraction network; Sub-step 404, downsampling, namely strengthening key features of ice and water differentiation through a deformable convolution layer and an attention mechanism; Sub-step 405, up-sampling, namely recovering the resolution of the image, and forming a classification probability map of the ice body and the water body; a substep 406 of converting the probability map into a binary mask with 0.5 as a threshold; In step 407, the ice sealing rate IC is calculated, and the formula: Ice sealing ratio ic=ice volume pixel number/(ice volume pixel number+water volume pixel number) ×100% Step 408, evaluating a result, namely comparing the predicted ice sealing rate with the identification accuracy ACC of the ice sealing rate of the reference icon, and evaluating the segmentation precision through the intersection ratio IoU of the predicted ice body region and the marked ice body region; And 5, outputting a visual image with the ice body marked white and the water body marked black and the value of the ice sealing rate IC.

Description

Intelligent recognition method for ice water characteristics and ice sealing rate under low-contrast scene Technical Field The invention relates to an intelligent recognition method for ice water characteristics and ice sealing rate under a low-contrast scene, which is a hydrologic calculation method and is an acquisition method for channel hydrologic data. Background There is a risk of ice disasters in river channels in alpine regions. The Jiang Hequ warehouse often generates ice flow in the initial icing stage or the unfreezing stage, a large amount of ice flow easily forms ice dams to block upstream water flow, and once the ice dams collapse, flood disasters can be formed. The ice sealing rate is an important parameter for influencing the loss of heat of a water body to ice and evaluating the risk of ice disaster. Through accurate ice sealing rate data, the ice flood disaster can be predicted efficiently and accurately, and the method is very important for disaster prevention and reduction. The definition of the ice sealing rate is quite simple, and the ice area on the water surface accounts for the ratio of the total area of ice and water. This data, while seemingly simple, is not readily available in a river and canal pool. The surface ice in the river in the cold region can last for more than 1000 kilometers, and the ice is continuously collided, piled and extruded in the flowing process, so that the process is quite complex. Early ice sealing rate monitoring relies on visual inspection, and has various problems such as large error, low frequency, incapability of verification and the like. In recent years, with the appearance of high-definition videos and the improvement of image recognition technology, the ice sealing rate monitoring method based on high-definition image recognition is fast in development, and various defects of traditional manual visual inspection are overcome to a great extent. Compared with river channels, the artificial water delivery engineering has the characteristics of low ice and water contrast, low water quality and the like due to the small change of ice and water boundaries and hydrodynamic conditions, and the ice and water contrast is difficult to distinguish on images, so that the error of the ice sealing rate identification method based on video images is obviously increased, the prediction capability of ice flood disasters is affected, and the problem is particularly prominent in large-scale water delivery engineering. How to accurately obtain the ice sealing rate in a water delivery channel is a problem to be solved. Disclosure of Invention In order to overcome the problems in the prior art, the invention provides an intelligent recognition method for ice water characteristics and ice sealing rate under a low-contrast scene. The method provides an intelligent ice sealing rate recognition algorithm based on a deformable convolutional neural network, and the algorithm introduces a deformable convolutional layer and can adaptively adjust the sampling position of a convolutional kernel so as to capture complex ice and water characteristics under a low-contrast scene more accurately. The invention aims to realize the intelligent identification method of the ice water characteristics and the ice sealing rate under a low-contrast scene, which comprises the following steps: The method comprises the steps of 1, acquiring video data of a monitoring camera, setting a time interval threshold t, intercepting an image frame from a read video at the threshold t, and then performing image format conversion and normalization size preprocessing; The method comprises the steps of marking data, finely dividing ice body and water body areas, generating a single-channel gray scale reference image, dividing the data, and classifying images in the data into a training set, a verification set and a test set according to the proportion of 75%, 20% and 5%; step 3, model construction and training, which comprises model construction and model training: (1) Model construction: The method comprises the steps of (1.1) extracting image features, namely extracting ice and water features in parallel by using a deep and shallow double-branch extraction network before a downsampling stage based on a U-Net model, sequentially inputting images into a conventional convolution layer with 3 x 3 convolution kernel sizes to realize feature information extraction, focusing on the whole outline of ice and water in the images by using a deep branch network, embedding an attention mechanism module behind each convolution layer by using 2 conventional convolution layers with 3 x 3 convolution kernel sizes, wherein the attention mechanism module comprises a channel attention mechanism for excavating important features of the ice and the water in a color channel dimension and a spatial attention mechanism for identifying positions of the important features in the images; The operation of the channel attention mechani