Search

CN-121998897-A - Method and system for detecting surface defects of railway precast beam

CN121998897ACN 121998897 ACN121998897 ACN 121998897ACN-121998897-A

Abstract

The invention relates to a method and a system for detecting surface defects of a railway precast beam, which relate to the technical field of precast beam detection, wherein the method for detecting the surface defects of the railway precast beam comprises the steps of image acquisition, image processing, image detection, output and the like, and the system for detecting the surface defects of the railway precast beam comprises an acquisition module, a processing module, a detection module and an output module. Upon detection of a crack, the channels associated with the "linear structure" are automatically enhanced, and the channels associated with the "uniform color patch" are suppressed. The system can intelligently ignore interference such as water drop reflection, shadow and the like which is different from the response of the defect characteristic channel, and the false alarm rate is reduced from about 18% to 2.6% of the traditional method. A large number of unnecessary reinspectors are reduced.

Inventors

  • ZHANG FANG
  • LIN SHICHEN
  • YUAN JUNJIE

Assignees

  • 西南交通大学

Dates

Publication Date
20260508
Application Date
20251208

Claims (9)

  1. 1. A method for detecting surface defects of a railway precast beam is characterized by comprising the following steps: the image acquisition is used for acquiring a digital image of the surface of the railway precast beam; An image process for normalizing and size normalizing the digital image to a fixed size; The image detection comprises the following steps: a feature extraction backbone network, which is an improved ResNet-34 network with a channel attention mechanism embedded in the residual block; The feature pyramid network is connected with the feature extraction backbone network and is used for merging and outputting a multi-scale feature map; The detection head is connected with the characteristic pyramid network and is used for predicting the category and position boundary box of the defect based on the multi-scale characteristic map; And outputting, namely generating classification and positioning information of the defects based on the multi-scale feature map output by the feature pyramid network.
  2. 2. A method for detecting surface defects of a railway precast beam according to claim 1, wherein the embedded channel attention mechanism is a Squeeze-and-specification module which is arranged after the last 1X 1 convolution of a main branch in a residual block and before summing with a shortcut connection branch.
  3. 3. The method for detecting the surface defects of the railway precast beam according to claim 2, wherein the operation of the Squeeze-and-specification module comprises the following steps: Squeeze operation to pool input feature graphs by global averaging Is used for spatial dimension compression of channel feature descriptors Wherein the calculation model of the c-th component is: ; Where u= { u 1 ,u 2 ,.........u C } is the input feature map, u c is the two-dimensional matrix of the c-th channel, the size is h×w, and z c is the global feature description of the c-th channel; The specification operation, namely inputting the channel characteristic descriptor z into a full-connection layer with a two-layer bottleneck structure, and firstly, passing through a weight matrix Reducing the dimension, activating by ReLU, and passing through weight matrix Restoring dimension, and finally generating weight vector of each channel through Sigmoid activation function ; Reweight operation of weighting the weight vector And original characteristic diagram Multiplying channel by channel, waiting for a calibrated feature map 。
  4. 4. The method for detecting the surface defects of the railway precast beam according to claim 3, wherein the construction process of the characteristic pyramid network comprises the following steps: The bottom-up path selects the output characteristic diagrams from the second stage to the fifth stage of the improved ResNet-34 network, and the output characteristic diagrams are marked as { C2, C3, C4, C5}; A top-down path, up-sampling 2 times from C5, and adding element by element with the next layer of feature map { C4, C3, C2} which is down-scaled to 256 channels through 1X 1 convolution; After each addition, a 3 x3 convolution layer is used to generate the output layers { P2, P3, P4, P5} of the feature pyramid, respectively.
  5. 5. A method for detecting surface defects of a railway precast beam according to claim 3, wherein the detection heads are respectively connected to each output layer { P2, P3, P4, P5} of the characteristic pyramid network, so that defect prediction is respectively carried out on characteristic diagrams of different scales.
  6. 6. The method for detecting the surface defects of the railway precast beam according to claim 5, wherein the step of acquiring the images is characterized in that a multi-angle synchronous shooting strategy is adopted, and the method comprises the step of synchronously acquiring the images of different curved surfaces of the beam body by using industrial cameras arranged on the precast Liang Zhengmian, the inclined side surfaces and the side surfaces.
  7. 7. The method for detecting surface defects of a railway precast beam according to claim 6, wherein the size normalization performed in the image processing step is performed with a fixed size of 224 pixels by 224 pixels or 448 pixels by 448 pixels.
  8. 8. The method for detecting surface defects of a railway precast beam according to claim 7, further comprising a data fusion interface, wherein the data fusion interface is used for receiving data from ultrasonic detection equipment and/or infrared thermal imaging equipment, and performing collaborative analysis and comprehensive diagnosis on surface defect information and internal defect information output by the core detection network module.
  9. 9. A surface defect detection system for a railway precast beam, which uses claim 1 4, Is characterized by comprising an acquisition module, a processing module, a detection module and an output module; the output end of the acquisition module is electrically connected with the input end of the processing module and is used for acquiring images; The output end of the processing module is electrically connected with the input end of the detection module and is used for processing the image set acquired by the acquisition module; The output end of the detection module is electrically connected with the input end of the output module and is used for detecting the defects of the precast beam; And the output module is used for outputting the defect type of the precast beam.

Description

Method and system for detecting surface defects of railway precast beam Technical Field The invention relates to the technical field of precast beam detection, in particular to a method and a system for detecting surface defects of a railway precast beam. Background The railway precast beam is a concrete beam component which is prefabricated and molded in a factory or a precast yard, is transported to a construction site for assembly and erection, is used for an industrialized construction mode of a railway bridge structure, and can generate defects such as pitting surface, bubbles, cracks, honeycomb and the like on the surface of the precast beam after pouring. At present, when the precast beam is detected, the detection and the machine vision detection are adopted, but the manual detection and the machine vision detection have a plurality of short plates: The manual detection limitation is that the detection is influenced by subjective experience and fatigue degree depending on human eyes, the omission factor is as high as 10% -30%, and the defect size (such as crack length and pit depth) cannot be quantified. The traditional machine vision short plate is based on fixed threshold segmentation, edge detection and other algorithms, has poor adaptability to beam bodies with illumination change and complex surface textures (such as concrete honeycomb and exposed steel bars), and is easy to misjudge. Therefore, there is a need for a method for detecting surface defects of a railway precast beam, which improves the accuracy of detecting the surface defects of the precast beam. Disclosure of Invention In order to improve the detection accuracy of the defects of the precast beam, the invention provides a method and a system for detecting the defects of the surface of the railway precast beam. In a first aspect, the present invention provides a method for detecting a surface defect of a railway precast beam, which adopts the following technical scheme: a method for detecting surface defects of a railway precast beam comprises the following steps: the image acquisition is used for acquiring a digital image of the surface of the railway precast beam; An image process for normalizing and size normalizing the digital image to a fixed size; The image detection comprises the following steps: a feature extraction backbone network, which is an improved ResNet-34 network with a channel attention mechanism embedded in the residual block; The feature pyramid network is connected with the feature extraction backbone network and is used for merging and outputting a multi-scale feature map; The detection head is connected with the characteristic pyramid network and is used for predicting the category and position boundary box of the defect based on the multi-scale characteristic map; And outputting, namely generating classification and positioning information of the defects based on the multi-scale feature map output by the feature pyramid network. Optionally, the embedded channel attention mechanism is a Squeeze-and-specification module, which is arranged after the last 1 x 1 convolution of the main branch in the residual block and before summing with the shortcut branch. Optionally, the operation of the Squeeze-and-specification module specifically includes the following steps: Squeeze operation to pool input feature graphs by global averaging Is used for spatial dimension compression of channel feature descriptorsWherein the calculation model of the c-th component is: ; Where u= { u 1,u2,.........uC } is the input feature map, u c is the two-dimensional matrix of the c-th channel, the size is h×w, and z c is the global feature description of the c-th channel; The specification operation, namely inputting the channel characteristic descriptor z into a full-connection layer with a two-layer bottleneck structure, and firstly, passing through a weight matrix Reducing the dimension, activating by ReLU, and passing through weight matrixRestoring dimension, and finally generating weight vector of each channel through Sigmoid activation function; Reweight operation of weighting the weight vectorAnd original characteristic diagramMultiplying channel by channel, waiting for a calibrated feature map。 Optionally, the construction process of the feature pyramid network includes the following steps: The bottom-up path selects the output characteristic diagrams from the second stage to the fifth stage of the improved ResNet-34 network, and the output characteristic diagrams are marked as { C2, C3, C4, C5}; A top-down path, up-sampling 2 times from C5, and adding element by element with the next layer of feature map { C4, C3, C2} which is down-scaled to 256 channels through 1X 1 convolution; After each addition, a 3 x3 convolution layer is used to generate the output layers { P2, P3, P4, P5} of the feature pyramid, respectively. Optionally, the detection head is connected to each output layer { P2, P3, P4, P5} of the feature pyramid network