CN-121981993-A - X-ray weld defect detection method based on structural degradation correction and parallel response recombination

Abstract

The invention provides an X-ray weld defect detection method based on structural degradation correction and parallel response recombination, which is applied to a single-stage target detection network, and is used for inhibiting structural information degradation caused by scale change and resampling operation by introducing a structural constraint mechanism which does not depend on a learnable parameter in a characteristic transmission process and reconstructing a response organization mode in the same characteristic level, correcting transmission characteristics by adopting a fixed differential constraint operation in a characteristic transmission process from deep to shallow, and realizing parallel response composition of multiple space coverage characteristics in the same level under the condition of not increasing network depth and characteristic level.

Inventors

• LUO RENZE
• LIANG XIAOYAN
• LIU QI
• LI JINTAO
• WANG GUANSHENG
• CHEN XINGTING
• LI ZHIQI
• ZHANG LUNFENG
• QIU JUNJIE
• LIU JIA

Assignees

• 西南石油大学

Dates

Publication Date

20260505

Application Date

20260124

Claims (1)

1. 1. The X-ray weld defect detection method based on structural degradation correction and parallel response recombination is characterized by comprising the following steps of: Step 1, constructing an X-ray weld defect data set F, , C, H, W respectively represents the channel number, the height and the width of the input feature map, and randomly divides the training set, the verification set and the test set according to proportion by using a script; step 2, constructing a target detection network model for detecting the X-ray weld defects, which mainly comprises the following steps: Step 2.1, constructing a backbone network for extracting characteristics of an input X-ray welding line image: The method comprises the steps of performing step-by-step downsampling and feature modeling on an input X-ray weld image by a main network through a layered feature extraction structure based on a convolutional neural network, performing feature extraction on the image by a multi-layer convolutional module at the front part of the network, wherein the convolutional module adopts a convolutional kernel with the size of Convolution operation with step length of 2, namely, gradually reducing the spatial resolution of the feature map and simultaneously improving the channel dimension so as to obtain feature representations with different scales; Feature enhancement units are introduced between adjacent convolution layers, and features are modeled and transferred in a multi-convolution transformation and cross-layer feature connection mode, so that feature expression capacity is enhanced and feature propagation stability is improved on the premise of keeping the feature space scale unchanged; At the high-level stage of the backbone network, a spatial pyramid pooling module is introduced to carry out multiscale receptive field aggregation on high-level features, and further, a feature enhancement module is used for carrying out self-adaptive enhancement on high-level semantic information, so that multi-level feature representation containing rich spatial structure information and high-level semantic information is output while the compactness of the network structure is maintained, and input is provided for a subsequent feature fusion neck network; Step 2.2, constructing a feature fusion neck network formed by a structural degradation correction module and a parallel response recombination module in a cooperative manner: The feature fusion neck network is positioned between the main network and the detection head, the input of the feature fusion neck network is a multi-scale feature map output by the main network, and the output of the feature fusion neck network is a multi-scale fusion feature map for the detection head; Step 2.2.1, constructing a structural degradation correction module: the structural degradation correction module is a zero-parameter module which does not contain a learnable parameter and is derived from shallow detail features with higher spatial resolution in a main network And deep semantic features derived from lower spatial resolution and stronger semantic information in a backbone network or neck network The shallow detail features are noted as: ; the deep semantic features are noted as: ; Wherein, the Respectively represents the number, the height and the width of the channels of the feature map, and , ; The implementation process of the structural degradation correction module comprises the following steps: for shallow detail features Applying a fixed form of low-pass constraint operator To obtain low frequency structural components : ; By shallow detail features And low frequency structural components Construction of structural residual information from differences between : ; For deep semantic features Upsampling by bilinear interpolation To make its spatial resolution Consistent, deep semantic features are obtained : ; Completion of structural residual information by channel alignment strategy without introducing any learnable parameters And deep semantic features after upsampling The channel alignment strategy comprises at least one mode of expanding the number of channels through a channel replication mode, compressing the number of channels through a channel clipping mode and dividing the channels into a plurality of groups and summing the channels in each group to obtain the target number of channels, wherein the structure correction characteristics after channel alignment are recorded as follows: ; structural correction features after alignment of the channels Deep semantic features injected after upsampling Obtaining structural correction features : ; The final output feature is obtained by correcting the structure With original shallow detail features And the method is obtained by splicing the dimensions of the channel dimension, namely: ; Wherein, the As an input feature to the subsequent parallel response reorganization module, Representing feature combination operators that stitch input features in the channel dimension to form a joint feature representation that contains semantic information and structural detail information For a fixed kernel form of a space smoothing operator, the kernel size and step size are characterized by shallow details For obtaining a stable low frequency response without introducing a learnable weight, the values of which include, but are not limited to, low pass filtering based on mean pooling, spatial smoothing based on a fixed convolution kernel; step 2.2.2, constructing a parallel response recombination module: The parallel response reorganization module is used for completing multi-scale space response modeling in the same feature level, and inputting features The method is characterized by comprising the following steps: ; and the parallel response construction is carried out according to the following structural flow: input features are first entered in the channel dimension Equally divided into two parts in the channel dimension: ; Wherein, the The said As a reserved branch, the method is used for maintaining the structural continuity of the original feature layer so as to avoid the damage to the original response mode in the parallel response recombination process; as input to parallel scale modeling branches; Branching the parallel scale modeling The channel dimension is further divided into four sub-branches, namely a micro-feature group, a mesoscopic feature group, a macro-feature group and a reserved group: ; Wherein, the channel number of each sub-branch is the same, and the spatial resolution is kept unchanged; applying convolution kernels of size to each sub-branch separately The depth separable convolution of (1) is characterized in that the micro-feature group is subjected to standard convolution treatment and is used for capturing micro point-like defect features, the mesoscopic feature group is subjected to cavity convolution treatment with the expansion rate of 2 and is used for capturing crack features, the macroscopic feature group is subjected to cavity convolution treatment with the expansion rate of 3 and is used for capturing background context information, and the reservation group is not subjected to convolution treatment and directly reserves input information; splicing the output characteristics of each sub-branch in the channel dimension, and reserving branches with the characteristics Splicing again, and obtaining output characteristics of the parallel response recombination module after subsequent convolution fusion, wherein the output characteristics are used as an input characteristic diagram of a subsequent detection head; step 3, a training model of a target detection network method for detecting the defects of the X-ray welding seams constructed based on the step 1 comprises the following steps: Step 3.1, setting training parameters of the network model, including batch size Rate of learning Optimizer, thread count Number of training wheels All parameters are optimal values obtained through multiple experimental comparison; Step 3.2, inputting the training set and the corresponding labeling data in the step 1 into the ray image weld defect detection model constructed in the step 2 for training, evaluating the model performance by using the verification set, and storing an optimal model weight file of a complete training period; And 4, adopting the optimal model weight stored in the step 3.2 to be used for the test set data in the step 1, outputting a visual detection result of the weld defects and evaluating the model detection performance.

Description

X-ray weld defect detection method based on structural degradation correction and parallel response recombination Technical Field The invention relates to the technical field of computer vision and industrial nondestructive detection, in particular to a defect detection method for an industrial pipeline welding seam X-ray image, and specifically relates to a target detection technology which corrects structural degradation in a target detection network characteristic propagation process and completes parallel response recombination in the same characteristic level. Background The welding quality of the pipeline is directly related to the safety and reliability of an industrial production system, and X-ray detection is used as a mainstream nondestructive detection means and is widely applied to detection and evaluation of internal defects of welding seams. With the development of deep learning technology, the automatic defect detection method based on the convolutional neural network gradually replaces the traditional manual film evaluation mode, wherein a single-stage target detection algorithm represented by the YOLO series is widely applied to the field of industrial defect detection due to high detection speed and flexible deployment. The existing X-ray weld defect detection method based on deep learning is mainly based on a single-stage target detection network, and the detection performance is improved by introducing an attention mechanism, multi-scale feature fusion or a leachable weight modulation module. For example, the patent CN120374557A, CN120747745A and CN121170754A each enhance the characterization capability of the model on defects of different scales by constructing a multi-scale feature pyramid, cross-level feature fusion or channel semantic weighting, while the patent CN113674247A, CN120807437a and CN121330550a modulate the feature response by introducing a channel or spatial attention mechanism, so as to enhance the identification effect on defects of low contrast. In addition, some of the prior art, such as CN120689287A, CN121095194a, improves detection stability from the perspective of response modulation or system level optimization by introducing geometric modulation, multi-modal fusion, or confidence weight modeling, etc. However, the above-described methods generally rely on the introduction of a learnable parameter or a multi-level feature stack structure, the technical emphasis of which is to enhance feature expression capability, without specific constraints on the problem of structural information degradation caused by resampling and scale change during top-down propagation of features. Meanwhile, the multi-scale step-by-step stacking structure often brings additional parameter redundancy and calculation overhead, and is not beneficial to the requirements of industrial sites on real-time performance and lightweight deployment. Therefore, it is necessary to propose a new feature processing mechanism that suppresses the structure degradation accumulation problem from the viewpoint of feature propagation structure stability without introducing a complex structure and additional learnable parameters. Disclosure of Invention 1. The invention aims to: The invention aims to provide a target detection method for detecting the defects of an X-ray welding line, which is characterized in that a structural degradation correction mechanism is introduced in the characteristic propagation process, parallel response recombination is completed in the same characteristic level, and the detection precision and stability of the defects of the welding line with low contrast and multiple forms are improved while the calculation complexity of a model is reduced. 2. The technical scheme is as follows: The X-ray weld defect detection method based on structural degradation correction and parallel response recombination is characterized by comprising the following steps of: Step 1, an X-ray weld defect detection method based on structural degradation correction and parallel response recombination is characterized by comprising the following steps: Step 1, constructing an X-ray weld defect data set F, ,C, H, W respectively represents the channel number, the height and the width of the input feature map, and randomly divides the training set, the verification set and the test set according to proportion by using a script; step 2, constructing a target detection network model for detecting the X-ray weld defects, which mainly comprises the following steps: Step 2.1, constructing a backbone network for extracting characteristics of an input X-ray welding line image: The method comprises the steps of performing step-by-step downsampling and feature modeling on an input X-ray weld image by a main network through a layered feature extraction structure based on a convolutional neural network, performing feature extraction on the image by a multi-layer convolutional module at the front part of the network, wherein