CN-121999349-A - Underwater dam defect light-weight segmentation method for underwater robot

CN121999349ACN 121999349 ACN121999349 ACN 121999349ACN-121999349-A

Abstract

The invention discloses an underwater dam defect light-weight segmentation method for an underwater robot, and relates to the technical field of image processing methods. The method comprises the steps of obtaining a dam underwater defect data set, preprocessing the data set, constructing LPFNet detection models for segmenting an underwater defect image, training the LPFNet detection models by using training sets and verification sets to obtain a trained LPFNet detection model, inputting test set pictures into the trained LPFNet detection model, checking generalization capability of the model, obtaining underwater defect pixel level results by morphological operation of defect segmentation results output by the LPFNet detection model, and converting the dam underwater defect pixels into actual sizes by combining a camera imaging principle to obtain real quantification results of the underwater defects. The method can realize real-time and quantitative identification of the defects of the underwater dam and improve the accuracy of detecting the defects of the underwater dam.

Inventors

KANG FEI
Tang Huadu
WU YINGRUI
LI JUNJIE
LI HONGQUAN
LIANG CHENXI

Assignees

大连理工大学

Dates

Publication Date: 20260508
Application Date: 20260104

Claims (10)

1. An underwater dam defect light-weight segmentation method facing an underwater robot is characterized by comprising the following steps of: acquiring a dam underwater defect image data set and preprocessing the data set; constructing LPFNet detection models for segmenting the underwater defect images; Training LPFNet the detection model by using the training set and the verification set to obtain a LPFNet detection model after training; Inputting the test set picture into a LPFNet detection model after training, and checking the generalization capability of the model; carrying out morphological operation on a defect segmentation result output by the LPFNet detection model to obtain an underwater defect pixel level result; and converting the pixels of the underwater defects of the dam into actual sizes by combining with a camera imaging principle to obtain the actual quantification result of the underwater defects.
2. The underwater robot-oriented underwater dam defect lightweight segmentation method as set forth in claim 1, wherein the method of acquiring and preprocessing a dam underwater defect dataset comprises the steps of: The defect image data set comprises crack and peeling image data, the image data is used for shooting the underwater surface of the dam by using an underwater robot, an underwater high-definition image is acquired by adopting an optical sensor, the preprocessing comprises scaling original image pixels to 256×256 and image enhancement, meanwhile, the three types of cracks, peeling and backgrounds are marked manually, and the data set is divided into a training set, a verification set and a test set according to the proportion of 7:2:1 after marking is finished.
3. The underwater robot-oriented underwater dam defect lightweight segmentation method as set forth in claim 1, wherein the constructing LPFNet detection model for segmenting the underwater defect comprises: Improving an LDB module in an encoder of the LETNet model, and replacing the LDB module in the LETNet model with a ALDB module; the ET module of the backbone network of the LETNet model was modified and HLPF module was used to replace the ET module in the LETNet model.
4. The underwater dam defect light-weight division method for the underwater robot according to claim 3, wherein the processing method of ALDB modules comprises the following steps: The input features are compressed to C/2 through 1X 1 Conv, the features after dimension reduction are sequentially subjected to two decomposition convolutions of 3X 1 and 1X 3 Conv to extract defect local features in vertical and horizontal directions, the local features enter into cavity depth separable convolution branches, the 3X 1 and 1X 3 DDConv with different cavity rates are used for expanding the receptive field, global dependence modeling capacity is enhanced, the cavity convolution outputs are then entered into AWSA modules to establish spatial position correlation, spatial feature response is improved, the branch outputs and dimension reduction features are added and then restored to original channels through 1X 1 Conv, finally, channel shuffling is carried out on restored features through ShuffleBlock, channels are divided and rearranged according to groups to establish information flow of different channels, and cross-layer residual fusion is achieved.
5. The underwater dam defect light-weight segmentation method for the underwater robot according to claim 4, wherein the AWSA module processing method comprises the following steps: Firstly, carrying out average pooling on input features in XY directions to obtain two one-dimensional feature graphs, re-expanding the two one-dimensional features to the original feature size, carrying out weighted fusion through learnable parameters to realize self-adaptive feature recalibration, generating a preliminary spatial attention map through Sigmod on the fused feature graphs, carrying out channel average on the attention map and 3X 3 Conv2d, and finally multiplying the attention map with the input features element by element to strengthen an underwater defect area and inhibit invalid features, wherein the operation formula of the AWSA module is as follows: ; ; ; ; ; Wherein, the Is an input feature that is used to determine the input, , Average pooling in horizontal and vertical directions respectively, , Is a weight that can be learned and the weight, Representing the multiplication by element, Placing the denominator as 0, Representing taking the average along the channel dimension.
6. The underwater dam defect light-weight segmentation method for the underwater robot according to claim 3, wherein the method comprises the following steps: The HLPF module adopts hierarchical feature complementation of the FPN-PAN module, and is assisted by the lightweight dynamic module LiteDPG Block and the HFB module, the FPN-PAN module receives three layers of underwater image semantic features from ALDB low, medium and high, the upper layer features are sampled and fused with the medium and low layer features step by step, and then reverse aggregation is carried out to transmit the low layer detail information back to the high layer features, so that the features of different layers are fully complemented.
7. The underwater dam defect light-weight segmentation method for the underwater robot as set forth in claim 6, wherein the LiteDPG Block structure processing method comprises the following steps: Firstly, channel compression is carried out through 1X 1 Conv, then the local space sensing capability of the model is enhanced through 3X 3 DConv, local feature cross-channel interaction is realized through ECA, DPG is introduced to further improve defect semantic features, the model can dynamically adjust feature distribution according to input features through multiplication branches and addition branches, finally, the model is restored to original dimensions through 1X 1 Conv, residual connection ensures feature consistency and gradient stability, and an operation formula of LiteDPG Block is as follows: ; ; ; ; ; Wherein, the Is an input feature that is used to determine the input, Is the output of the deep convolution, Is the output of the ECA, Is the output of the DPG and, Is a convolution module, comprising convolution, batchnorm, preLU operations; is a depth-separable convolution module, Is a 1-dimensional convolution operation.
8. The underwater dam defect light-weight segmentation method for the underwater robot according to claim 6, wherein the method comprises the following steps: The HFB module utilizes three scale features output by FPN-PAN, guides and weights middle layer features according to high-level semantic information, fuses the middle layer features with bottom layer features to realize effective integration of global semantics and local details, enhances the recognition capability of an enhanced model on underwater micro defects and complex backgrounds, extracts global context information through average pooling, then carries out channel mapping and normalization through 1X 1 Conv2d and Sigmod, multiplies the extracted global context information by middle layer features element by element, samples different features to the same scale for splicing, carries out local information fusion through 3X 3 Conv, and obtains cross-layer mixed features, wherein the operation formula of the HFB module is as follows: ; ; Wherein, the , , Respectively high, medium and low layer features, Is an upsampling operation.
9. The underwater dam defect light-weight segmentation method for the underwater robot according to claim 1, wherein the method for obtaining the underwater defect pixel level result by morphological operation of the defect segmentation result output by the LPFNet detection model comprises the following steps: Firstly, carrying out binarization processing on a mask image according to the defect mask image which is output by a detection model and contains crack and peeling type information, then traversing the binarization image by adopting an 8-neighborhood connected detection algorithm, extracting independent connected domains of each segmentation result, and respectively calculating pixel-level quantization parameters aiming at the connected domains of different types; for the peeling defect, counting the total number of foreground pixels in the connected domain to obtain peeled pixel-level area parameters, wherein the calculation formula is as follows: ; Wherein, the Pixel level area representing a peeling defect; representing a connected domain region where the peeling defect is located; at coordinates representing a binarized mask image Pixel values of (2); for the crack defect, extracting a skeleton of a crack region, performing skeleton trimming treatment, and measuring the crack width by using an improved Euler method based on the trimmed skeleton, wherein the calculation formula is as follows: ; ; Wherein, the Is the total width of the crack at point x; The arc length of the track from the point x to the upper boundary; the arc length of the track from the point x to the lower boundary is shown, s and T are arc length parameters, ds and dt are integral step sizes, and T is a unit tangential vector field; Satisfy the Laplace equation Is a potential function of (2); The gradient as a potential function indicates the direction of extension of the crack width.
10. The underwater robot-oriented underwater dam defect light-weight segmentation method as set forth in claim 1, wherein the method for converting the dam underwater defect pixels into actual sizes by combining the camera imaging principle to obtain the actual quantification result of the underwater defect comprises the following steps: According to imaging parameters of an underwater robot carrying camera, a mapping relation between a pixel coordinate system and a world coordinate system is established, the actual physical width and the actual physical height represented by a single pixel in an image are calculated firstly by utilizing the field angle of the camera, the object distance during shooting and the image resolution, and the obtained pixel-level detection result is converted into an actual millimeter-level result based on the calculated single-pixel physical size, wherein the calculation formula is as follows: ; ; Wherein, the Representing the true physical area of the spall area; representing the true physical width of the fracture zone; Representing the actual physical width represented by each pixel in the image; Representing the actual physical height represented by each pixel in the image.

Description

Underwater dam defect light-weight segmentation method for underwater robot Technical Field The invention relates to the technical field of image processing methods, in particular to an underwater dam defect light-weight segmentation method for an underwater robot. Background The dam is an important infrastructure in hydraulic engineering and has important economic and social significance. The dam has important functions of hydroelectric power generation, flood control, agricultural irrigation and the like. The underwater part of the dam is subjected to multiple functions such as water flow flushing, water pressure, sediment erosion and the like for a long time, and the defects such as cracks, flaking and the like are easy to generate. These defects can seriously impair the strength, durability and overall stability of the structure. Therefore, timely and accurate detection and maintenance are critical to maintaining safe operation of the dam. The traditional dam underwater detection method mainly comprises manual diving inspection, water storage layer emptying inspection, embedded sensor monitoring and the like. The manual inspection is limited by depth and environment, the risk is high, the speed is low, and the detection result is influenced by subjective factors. Draining the reservoir not only wastes water resources, but may also alter the stress distribution of the dam structure. And the sensor may provide monitoring data to determine the location and status of the lesion. However, long-term underwater deployment has reliability problems, and effective defect detection is difficult to achieve. By combining a deep learning method, the underwater robot can identify defects such as cracks and flaking, acquire spatial position and boundary information of the defects, and realize quick decision and accurate maintenance. However, the underwater environment is complex, and the image degradation can be caused by factors such as illumination attenuation, water turbidity, noise interference, light refraction and the like, so that the difficulty in identifying the underwater defects is greatly increased. Based on the existing deep learning method, the method has limited expression in the underwater environment, and is difficult to realize reliable perception of fine defects such as underwater cracks and the like. In addition, the underwater robot is limited in computational resources, and large-scale network deployment real-time detection is difficult to deploy. Therefore, a high-precision, lightweight and highly robust depth model suitable for the underwater environment of the dam needs to be established to realize automatic and real-time underwater defect detection. Disclosure of Invention The invention aims to solve the technical problem of providing an underwater robot-oriented underwater dam defect light-weight segmentation method which can realize real-time and quantitative identification of the underwater dam defect and improve the detection accuracy of the underwater dam defect. In order to solve the technical problems, the technical scheme adopted by the invention is that the underwater dam defect light-weight segmentation method facing the underwater robot comprises the following steps: Acquiring a dam underwater defect data set and preprocessing the data set; constructing LPFNet detection models for segmenting the underwater defect images; Training LPFNet the detection model by using the training set and the verification set to obtain a LPFNet detection model after training; Inputting the test set picture into a LPFNet detection model after training, and checking the generalization capability of the model; carrying out morphological operation on a defect segmentation result output by the LPFNet detection model to obtain an underwater defect pixel level result; and converting the pixels of the underwater defects of the dam into actual sizes by combining with a camera imaging principle to obtain the actual quantification result of the underwater defects. The technical scheme has the beneficial effects that the method aims at complex underwater scenes of the dam, a LPFNet detection network is built on the basis of LETNet, multi-scale context information and an underwater defect characteristic enhancement mechanism are fused, and high-precision segmentation and real-time identification of defects such as underwater cracks and flaking are realized. Meanwhile, the network can be ensured to be deployed on the underwater robot in a lightweight way. The physical size of the underwater defect can be rapidly calculated through the parameters of the camera and the imaging principle, and the damage condition of the underwater defect of the dam can be quantified. Drawings The invention will be described in further detail with reference to the drawings and the detailed description. FIG. 1 is a main flow chart of a method according to an embodiment of the invention; FIG. 2 is a block diagram of LPFNet detection model