Research On Defect Detection Algorithm Of Photoelectric Components Based On Convolution Neural Network

As the industrial 4.0 era witnesses the rapid development of China's manufacturing industry,enterprises have set higher requirements for the real-time performance and quality of defect detection in production.Thus,it has become a hot topic for scholars to achieve defect detection automation through machine vision and deep learning.Photoelectric components serve as the fixed components of mobile phone cameras,and it is particularly significant to quickly detect the quality of this component and use manipulator to remove defective products out of the production line.However,in the production of photoelectric components,defects are still detected manually,which causes high labor cost,slow speed,and low degree of automation.In order to overcome the above-mentioned disadvantages,this paper integrates the existing traditional image processing and deep learning algorithms,and optimizes the model according to the characteristics of defective photoelectric components.The specific contents of this paper are as follows:Firstly,the production flow of photoelectric components is introduced,and the detection object of this subject is defined,and then the defect image acquisition system of photoelectric components is established.On such a basis,a large number of component image data are obtained by using the acquisition system,and then the data set and labels are made.Secondly,the model number and hole number detection model is constructed on the basis of Nano Det.By lightweight network,multi-scale information fusion and multi-branch prediction,Nano Det constructs a target detection model balancing both accuracy and real-time performance.As the number and size of model number and hole number are determined,the Backone,PAN(Path Aggregation Network)and detection head of Nano Det are optimized to reduce the complexity of the model and improve the real-time performance.The improved model ensures the accuracy and has better real-time performance in the detection task.Lastly,a burn defect detection model for lower die thread end is built based on traditional image processing and CNN.Aiming to reduce the interference of irrelevant elements and to receive more useful information,the burn picture of lower die thread end is preprocessed,mainly about grayscale,Gaussian filtering,binarization,Canny,etc.The position of the feature area is determined by contour information,and polar coordinates are transformed on the feature area.Then the transformed image is input into CNN so as to train the model,and the trained network can be employed to detect the defects of the component.The detection speed of the algorithm meets the requirements on burn data sets,and the accuracy rate reaches 96%.