Research On Defect Detection Technology Of New Material Flooring Based On Deep Leaning

The reduction of forest resources in the world has caused the domestic forestry to be limited,which restricted the development of the floor industry.In order to break through the bottleneck,and at the same time,keep up with the ‘Green’ in ‘Five Concepts For Development’ proposed by the Central Committee of the Communist Party of China(CPC),one of the flooring industry reform direction is to replace wood flooring with new material flooring.So,for improving the utilization rate of new material flooring,the defect detection of new material flooring becomes an important research topic.At present,the defect detection of new material floor often proceeds artificially with lower detect precision but longer time-consuming.At the same time there are too many first-line personnel,facing the machine safety hazard and the pollution such as dust,paint and noise.Aiming at the above problems,this paper studies a new material floor defect detection method,which process automatically,accurately and quickly,including Classification Task and Detection Task.The main research contents are as follows:(1)Set up the defect database of new material floorings.The whole object,defect type,detection difficulty and sample data are analyzed.The image acquisition system is designed.The combined sample data set was made for Classification Task,and the data enhancement and sample labeling were completed for Detection Task.(2)A defect classification method based on feature fusion convolutional neural network is proposed.After the experiment about perspectives of branch network infrastructure,feature fusion mode and feature fusion location,a Three-branch Feature Fusion Convolutional Neural Network--TFFCNN is designed,which takes Res Net-34 as the branch network infrastructure,Concatenation operation as the feature fusion mode,Conv5＿x as the feature fusion location,and embedded in the attention module of CBAM.And the accuracy of TFFCNN is 1.28%-2.86% higher than the traditional classification neural networks.(3)A defect detection method based on improved YOLOv5 is proposed.By embedding SE attention module and adding a detection head on the original YOLOv5 s model,the YOLOv5s-SE-4D model was formed.The model has the detection speed close to YOLOv5 s and the detection accuracy close to YOLOv5 l,but the number of parameters is much lower than YOLOv5 l.The experimental results show that the precision and real-time ability can achieve a high standard that meets industry requirements,meanwhile,the model takes up less memory and is easier to be embedded into devices with smaller memory.(4)A lightweight network model is proposed and applied to the Classification Task and Detection Task,respectively.First,the lightweight network model--FSConv and FSBneck module,is designed,which can "plug and play" to replace the primary convolutional module in the network.The subsequent experiments show that this module can approach or even exceed the accuracy of the baseline network while reducing a large quantity of parameters and Flops.At last,the lightweight network model is applied to the algorithm of Classification Task and Detection Task,which increases the flexibility by reducing parameters and Flops.(5)A software "New Material Floor Defect Detection System" was designed and implemented,which deploys the Classification Task and Detection Task.Although the image selection is processed manually at present,it provides an idea and framework for subsequent functional updates,so that the quality control personnel can process the quality test more conveniently and quickly.