Image Based Neural Network Method For Detecting Structural Surface Defects

Concrete structure and steel structure can be seen everywhere in life,they play a very important role in industrial production,transportation and construction engineering.These structures may have defects caused by accidents during manufacturing and fatigue during service.The traditional defect recognition methods based on human eyes and ears cannot meet the requirements of automation,high efficiency and high accuracy,so it is urgent to develop new structural defect recognition methods.In recent years,the image processing technology based on convolutional neural networks(CNN)has made a breakthrough.Using the cross knowledge of convolutional neural networks and image processing to detect structural surface defects is a research hotspot in the field of structural condition monitoring and health management.In this paper,the neural network recognition method of structural surface defects based on image is studied,the specific work contents are as follows:(1)Railway sleeper defect recognition based on edge detection enhanced convolution neural networkIn order to detect whether there are cracks in the sleeper image,we propose a two-stage sleeper crack recognition method based on edge detection and convolution neural network.In the first stage,3*3 neighborhood range algorithm is used for edge detection to search the possible crack area in the image,and a fixed threshold is used for binarization.Then,a series of mathematical morphology operations are used to eliminate the noise and noncrack targets in the edge detection results,and a binary image containing a small amount of noise,sewer,stains and other targets is obtained.In the second stage,a CNN model is built to further identify each target in the filtered edge detection results,and make the final judgment on whether the image contains cracks.Experiments show that the proposed 3*3 neighborhood range algorithm can detect cracks more clearly than Sobel and Canny algorithm,and the accuracy of edge detection results of CNN model classification can reach 95%.(2)Two stage method of convolution neural network initial inspection and Unet fine inspection for steel plate defect recognitionIn the railway sleeper defect recognition,we use the edge detection enhanced CNN model to accurately judge the sleeper surface defects,but we cannot judge the position,size and shape of the defects in the image.As a kind of convolutional neural network,Unet has been used in many real image segmentation tasks and achieved good segmentation results.It can recognize the defects in the structure surface image pixel by pixel,so as to get more information about the defects.In this paper,the CNN model is used to judge whether there are defects in the steel plate image,and the steel plate image with defects is further detected by the improved Unet model.The Unet model is composed of an encoder and a decoder.The encoder extracts feature with large receptive field through multiple down-sampling,and the decoder generates segmentation results through multiple up-sampling.Therefore,the features of large receptive field cannot directly affect the segmentation results of the model,which limits its performance.A new classification branch is introduced between the encoder output and the model output of Unet model to improve the performance of Unet segmentation of defect images by using the large receptive field features extracted by the encoder.In order to avoid the negative impact of sample imbalance,the defect classes with small number of samples are repeatedly used in model training.To sum up,this paper studies the structural surface defect recognition method based on neural network through two practical defect detection projects.In the detection of railway sleeper cracks,the crack edge detected by our proposed 3*3 neighborhood range algorithm is clearer than that detected by Sobel algorithm and Canny algorithm.The filtering method based on morphological operation can filter a large number of non crack targets,and the accuracy of CNN model for crack recognition reaches 95%.When detecting the defects on the surface of steel plate,we introduce a new classification branch between the encoder output and the model output of Unet to improve its performance,and reuse the defects with a small number of samples in training to effectively improve the sample imbalance problem.