Research On Concrete Surface Damage Detection And Recognition Based On Vision Under Complex Conditions

Concrete has a very important role in the construction of infrastructure such as residential facilities,transportation,and public facilities.However,concrete can be damaged for various reasons.Traditional manual inspection can be affected by the subjective factors of personnel,the detection of fluctuating accuracy and inefficiency.Now often use the automated detection method only for a single normal environment to detect,change the use of the environment will also make the detection accuracy decreased.Therefore,for the problem of concrete damage detection under complex conditions,this paper proposes a neural network for concrete damage detection based on an attention mechanism and builds a damage detection system around this network.The main content of the thesis is as follows.1.The thesis firstly describes the current status of concrete damage detection and details the background and significance of concrete damage detection.then the problems and challenges of concrete damage detection are presented;then the current situation of concrete damage detection at home and abroad is analyzed,and various methods of concrete damage based on various methods are reviewed;then the technical principle of deep learning used in this thesis is analyzed.It is then analyzed that the current mainstream deep learning-based concrete damage detection algorithm for concrete detection in complex environments still suffers from incomplete detection of concrete damage detail features,low detection accuracy,and oversized algorithm models that cannot be adapted to engineering needs.Therefore,to address this problem,this thesis proposes a lightweight convolutional neural network concrete damage detection algorithm based on the attention mechanism and lightweight convolutional method.2.The thesis first analyzes the causes of each type of concrete damage cracks,points out several factors that affect the generation of concrete damage cracks,and then gives the types of crack shapes under the combined influence of each factor.To construct a complete complex conditional crack dataset,the thesis uses field photography and online collection for each type of crack image collection,uses labelme software for crack labeling,and in order to increase the depth of the dataset,adopts image enhancement techniques to increase the number of images in the dataset,and prepares for the constructed neural network.3.In this thesis,concrete damage detection research is conducted using deep learning techniques that have become popular in recent years.The thesis proposes a lightweight convolutional neural network concrete damage detection algorithm based on attention mechanism,Attention-Crack(ATcrack).The algorithm uses an encoder-decoder architecture.To reduce the model size and increase the speed of model operations.The backbone of the ATcrack uses a lightweight network.ATCrack adds a channel attention module in the encoder stage to improve the network’s extraction of crack detail features in concrete damage images,and adds a spatial attention module in the decoder stage to improve the network’s ability to sense and segment the location of cracks in concrete damage images.After ATCrack completes its training on the dataset,it is compared with other methods in a comprehensive manner,using both qualitative and quantitative approaches.In the comparison tests,ATCrack achieves an accuracy of up to 0.68,a recall of up to 0.86,an F1 score of up to 0.75,and an average cross-merge ratio of up to 0.64,and has state-of–the-art in crack detection compared with the other algorithms.4.Finally,to simulate the actual engineering needs,this thesis establishes a concrete damage detection system that integrates the algorithm with concrete damage acquisition.The system is divided into an image acquisition platform and an image detection platform.Through this system,the concrete damage images collected from the acquisition platform can be sent to the detection platform for detection and analysis,and the damage identification of concrete damage images is completed in the detection platform and the results are returned to the cell phone.