Research On Image Crack Detection Algorithm On Deep Learning

With the rapid development of deep learning,semantic segmentation technology occupies a very important position in computer vision.Structures are prone to produce cracks with great harm during manufacturing,transportation and application.This is because the structure has been subjected to concentrated stress and alternating load for a long time.Cracks in the structure will directly affect production and life.Crack detection is one of the important links in the maintenance of concrete structures.It can directly evaluate the safety and durability of concrete structures.However,traditional crack detection methods are time-consuming and labor-intensive,and rely on human experience and technical level to judge.Therefore,automatic crack detection is very important to quickly and accurately detect and identify cracks on the road.Semantic segmentation can completely split the crack,and has good application value in engineering.Firstly,aiming at realizing online real-time segmentation of mobile devices as targets,the existing deep learning target detection algorithm is analyzed,and the structure and design method of crack detection network based on convolutional neural network are given.The crack image detection algorithm is implemented.To more realistically simulate the actual collection situation,a collection system based on omnidirectional mobile robots was built.Secondly,on the premise of ensuring the accuracy of the image crack segmentation,improve the real-time segmentation speed;focus on the deep learning DeepLab V3 + algorithm,which directly uses the DeepLab V3 algorithm as the encoding module.The feature map output in the DeepLab V3 algorithm is 256 channels,of which Contains rich semantic information.By controlling the expansion rate of the atrous convolution in the network,a feature map with arbitrary resolution is output,thereby balancing the running time and accuracy.Thirdly,for the crack image(the water under the bridge is very heavy)contains water mist interference resulting in low crack detection accuracy,the dark channel defog algorithm is used to defog the picture,enhance the clarity of the image in a complex environment,and improve the image crack Of segmentation accuracy.Then,the DeepLab V3 + algorithm in deep learning is used for crack semantic segmentation.When the original algorithm model is used to split concrete cracks,because the semantic information is not strong enough,the recognition effect of concrete cracks is not ideal,and the accuracy is not high enough.DeepLab V3 + image crack detection algorithm.Using the smaller Resnet-101 basic network of FLOPs(Floating point operations),the improved DeepLab V3 + algorithm is used for crack semantic segmentation;at the same time,the positive and negative samples of the crack image are uneven,and the negative samples account for more,resulting in a very slow loss function decline,and the accuracy of segmentation decreases,using Focal Loss instead of Cross Entropy Loss to reduce the impact of negative samples on the model,so that the speed of Loss reduction is increased.Finally,an image crack detection system based on DeepLab V3 + is designed and implemented.The experimental results show that when using the improved DeepLab V3 + to detect concrete cracks,the detection accuracy is improved,FWIo U is increased by 3.86%,and MIo U is increased by 14.4%,the real-time performance meets the requirements of the crack detection system.