Semantic Segmentation Method For Concrete Crack Images Based On Deep Learning

Cracks are a significant flaw in the inspection of civil engineering infrastructure,and in recent years,the demand for effective crack detection methods has increased significantly.With the development of image acquisition equipment and breakthroughs in digital image technology,taking pictures of structures directly by special equipment equipped with high-definition cameras and using deep learning methods to detect crack types,locations,and widths in real time is a trend for future automated crack detection.In response to the problem of small percentage due to fine and small features of crack parts in crack images,this thesis conducts a lot of experiments on the basic image semantic segmentation structure and proposes a deep learning-based semantic segmentation method for concrete crack images,and further makes optimization of the model for real-time problems.The experiments demonstrate that the method can meet the basic real-time detection requirements even in the edge computing platform with low performance,and effectively circumvent the possible time delay brought by cloud computing.The main research contents and contributions of this thesis are divided into the following points.(1)To address the problem that the existing concrete crack dataset is not representative and the quality is uneven,this thesis takes concrete structure crack image samples by handheld devices.The processed crack images are labeled pixel by pixel using Label Img Plus tool,and finally a concrete crack dataset containing 2000 crack images and2000 masks is constructed.In addition,to enhance the convincing power of the proposed model,a public dataset of bridge structure cracks containing other similar materials is introduced for secondary experiments.This dataset also contains 2000 crack images and2000 masks.Finally,a new dataset is formed by fusing the dataset of this thesis with the publicly available dataset,with the aim of enhancing the robustness of the subsequent model after training with the dataset.(2)A full convolutional crack image semantic segmentation method based on an encoder-decoder structure is proposed to address the problem that concrete crack pixels are difficult to segment and easy to miss features.Firstly,an encoder-decoder structure using a residual structure is proposed for extracting and recovering crack features based on existing methods.Secondly,the effect of the basic residual blocks on the network performance is investigated to find the residual block parameters that are relatively suitable for crack segmentation.In addition,the effect of network width on network performance is investigated and used to determine the feature dimension transformation strategy of this model.Finally,a loss function that takes into account the focal loss and the similarity of different samples is proposed to improve the problem that the crack accuracy is masked due to the disproportion between crack pixels and background pixels.These studies are mainly aimed at filtering out the parts of different models that are suitable for semantic segmentation of crack images and solving the accuracy problem of the models.The main accuracy index of the final obtained model reached 87.00%.(3)To solve the problem of low speed of current concrete crack image semantic segmentation methods,an optimized full convolutional crack image semantic segmentation method based on the encoder-decoder structure is proposed.Firstly,the network depth of the basic encoder-decoder structure is reduced,and the 25-layer network is reduced to 19 layers,considerably reducing the amount of network parameters.Secondly,by employing standard convolution rather than the more time-consuming omni-dimensional dynamic convolution,the network’s reaction time is shortened.Finally,an enhancement module is inserted in the middle of the encoder-decoder structure to strengthen the connection of contextual features.The optimized model outperforms other networks in terms of accuracy and achieves several times the speed of other models,reaching 133 frames in the experimental setting of this thesis.(4)The feasibility of this thesis’ s approach is demonstrated by converting the trained model into an open neural network switching model and then performing a real detection task using an edge computing platform.Finally,the performance of the platform under different input sizes is summarized and the relevant deployment suggestions are given for the semantic segmentation task of concrete crack images using mobile devices.