Research On Surface Defect Detection Algorithm Based On Convolution Neural Network

With the development of modern industry,product quality testing has become a necessary step in the production process,and the detection of surface defects is an important part.The traditional surface defect detection using artificial visual method has the disadvantages of poor stability,long time consumption,high labor costs.Because the machine vision detection technology has the advantages of non-contact,stable and reliable,fast and efficient,high degree of automation,it has been widely recognized in the product surface defect detection.In this paper,the button surface detection algorithm is studied and completed the following works:(1)A new algorithm for the detection of button surface defects based on convolutional neural network is proposed.The algorithm mainly uses surface image extractiong,target surface area normalization and image sharpening for preprocessing,uses convolutional neural network to get the results,avoids the process of manual image feature extraction in traditional defect detection algorithm based on pattern recognition.(2)The different network parameters(network configuration,network layer number,convolutional network characteristic graph number,convolution kernel size),network function(activation function,sampling function)and Dropout layer for defect detection of convolutional neural network is examined.The results show that Pyramid double-layer structure,ReLu function and Dropout technology are beneficial to the improvement of network performance.In this paper,the button is taken as the research object,the image of the button surface is collected as the sample set,and the network is verified.Finally,a convolutional neural network model is designed to detect the defects on the button surface.(3)The algorithm is transplanted to the DSP smart camera and optimized.In this paper,the algorithm is tested on a set of 1178 images.The results show that the algorithm is effective and applicable.The experimental results show that the proposed algorithm has the advantages of no need of manual extraction of features,and the neural network model for the button surface can achieve a correct rate of 96.3% and 99.1% for the scratch defect and uneven.The algorithm is successfully transplanted to the DSP smart camera,and the running speed of 326ms/ frame is realized.