Research On Surface Defect Detection Technology Based On Embedded Machine-Vision System

In recent years,industrial quality inspection has become an increasingly important part of industrial production.Manual detection is still the main way of surface detection leading to low detection efficiency and high miss rate.With the gradual sound of image processing theory and the rapid development of hardware processor performance,machine vision technology has become a feasible and efficient solution.The surface is complex and the defects are Various.They are the problems that need to be solved in our research.The accurate and real-time detection affects both the efficiency and quality of the whole production process.This paper takes buttons as the research object.We pro an algorithm to solve the button surface defect detection task and improve the adaptability and robustness of the algorithm.The algorithm aims to solve the over-fitting issue caused by the imbalance of training samples.Finally,the algorithm is transplanted and optimized on the embedded platform to realize online detection.According to the researches of button samples,we find that the features of positive buttons are similar,while the negatives are various and different from positives.Therefore,we train the siamese network to obtain a mapping function,so that in the feature space,the positive samples are clustered and the defective products are scattered around the clusters after the feature mapping.Then a hypersphere is constructed to classify the features of samples.This paper designs a new loss function to guide and accelerate the training of siamese networks.In this task,comparing the famous triple loss function,the features extracted by our methods have higher cohesion and low coupling distribution.The model of this paper mainly focuses on the distribution of positive samples,therefore it can effectively suppress the adverse effects caused by the extreme imbalance of training samples.Experiments show that the method are efficiency to various kinds of buttons and defects,with the accuracy rate is over 98%.The accuracy for detecting low contrast defects and small defects is more than 95%.The algorithm is more versatile.In the case of unbalanced training data,the recognition rate of the algorithm is more than 94%.In order to realize online detection,we optimize our algorithms according to the DSP architecture.The optimized algorithm achieves a detection speed of 6 fps in a 594 MHz frequency DSP,and the speed is improved by 33.7 times,meeting the requirements of industrial online detection.