Research On Surface Crack Image Processing And Recogniti On Technology Based On Deep Learning

Buildings and products appearance and life are linked in infrastructure construction and production operations.Automatic surface identification is an important means to ensure production safety and product quality.Automatic recognition using computer vision technology is highly efficient and low cost,and is the key of future research and development.This paper takes pavement cracks containing multiple background textures and noises as the main research object,and combines the theory and algorithm of deep learning to discuss and study step by step the problems involved in the process of crack recognition,such as classification and segmentation accuracy,the difference between static and dynamic recognition,multiple background noises and unbalanced and insufficient data samples.First,this paper found that the crack identification model is more dependent and sensitive to shallow edge features.Then this paper constructed a global multiplexing structure of original features,which connects the shallow and deep layers of the network across layers,thereby improving the model’s utilization of edge information.At the same time,the structure can also activate inactive neurons in the deep layer,which helps the model to backpropagate.The proposed structure was equipped with backbone networks,and the experiments were conducted on three crack picture-level classification datasets to verify its superior performance.Second,this paper proposed a coder-decoder network based on feature fusion by using edge information to compensate for the lack of segmentation details in pixel-level crack recognition,obtaining more accurate information about the category of missing pixel points.Based on the feature fusion idea of channel stacking and late fusion,the original feature global multiplexing structure was transposed into the encoder-decoder network,and the information of the encoder module was incorporated into the decoder module across layers,so that the gradients of the encoder module can be transmitted to the decoder module of the same scale in an orderly manner.Segmentation experiments shown that the proposed method could obtain more accurate edge contours.Finally,using the time series information in the video stream data,a temporal recurrencebased MN-GRU network is proposed for the pixel-level identification of crack video streams.In this method,the lightweight network Mobile Net V3 was adopted as the encoding network for feature extraction.the GRU memory unit constitutes the decoding network for learning the temporal coherence between video frames.To verify the effectiveness of the method,the pavement crack video dataset Crack Video SUT was constructed by our own collections,and the experimental results showed that MN-GRU is better than the commonly used networks in terms of discrimination Accuracy.