Wall Defect Detection System Research Based On Lightweight Neural Network

With the development of society,the cost of manual labor continues to rise,and the technology continues to improve.The realization of mechanization and automation of interior decoration is the future trend.In the field of interior decoration,wall repair is one of the important tasks.For the automatic decoration robot,the premise of completing the decoration work needs to complete the detection of wall defects and the positioning of wall defects.In this thesis,deep learning and SLAM（Simultaneous localization and mapping）are used to accomplish these two tasks under the condition of low computing power.In the task of wall defect detection,it is necessary to realize the identification and classification of wall defects.With the help of the classification of the wall defects,the robot can complete the wall repair.The detection of wall defects can be basically realized through the object detection algorithm in the field of deep learning.However,the basic algorithm model requires high-power computing equipment.It not only requires higher cost,but also requires a larger space volume.These factors will limit the realization of decoration robots.In order to overcome the problem of high computing power,this thesis uses the technology of lightweight network model to compress the network model and accelerate model inference,thereby greatly reducing the computing power required by the network model.We adopt a scheme optimized on a small network module.On the basis of small network modules,16-bit and 8-bit quantization models are used to compress models and accelerate inference.Binary Networks are used to accelerate inference.Compact network design is used to reduce model parameters,thereby compressing the model.With the above network lightweight optimization design,the wall defect detection module can be running on embedded devices with low computing power.In the task of locating wall defects,it is necessary to build an indoor 3D（threedimensional）map,and mark the location of wall defects on the map.An end-to-end visual SLAM system can build indoor 3D sparse maps.However,these maps are composed of non-semantic feature points,and no wall defects are marked.In this thesis,by inputting the defect detection box and classification information output by the wall defect detection module into the visual SLAM front-end,the relationship between the defect point and the feature point set in the detection box is established..Estimating the 3D coordinates of defect points through feature point sets.The Hungarian allocation algorithm is used to associate the defect points between frames.Finally,the feature point set is registered into the3 D map as a map point.In this way,a 3D map with wall defect semantics is constructed,and a semantic SLAM system is realized.Through these two prerequisite tasks,the decoration robot can move to the wall defect position through the map,and implement different wall repair operations according to the defect type.