The Image Defect Detection Algorithm Based On Normal Sample Learning

As the deep learning technology has developed constantly over the years,the penetration of artificial intelligence into a wide range of industrial production fields can be seen.Machine vision-based object surface defect detection is urgently needed in actual industrial production.However,it is difficult to collect a large quantity and complete variety of defect samples due to the small probability and randomness of defect samples.It is difficult to apply the supervised learning method due to the imbalance of positive and negative samples as well as the changeable and unpredictable appearance of defects.In the meantime,it needs to consume huge human resources for the pixel level annotation of data sets.The thesis investigates and optimizes the use of knowledge distillation for computer vision defect detection.The basic idea is to make the application of the pre-trained teacher network and take advantage of the normal samples to teach students to study the output feature representation of the normal samples.During the test,the defect area can be detected and located by comparing the differences between the student network and the teacher network.The following is the main work of this thesis:(1)A distillation model algorithm for image defect detection with the introduction of an attention mechanism is proposed.Improvement is made based on the image defect detection algorithm of the distillation model in the paper,by introducing the attention mechanism into the model and optimizing the expressiveness of the network model based on the visual attention mechanism.Results of experiments show that accuracy in defect detection is improved by the improved model.(2)The loss function of the distillation model image defect detection algorithm is optimized.A loss function for the image defect detection algorithm applied to the distillation model is designed in the paper based on the idea that the SSIM((structural similarity index))algorithm calculates the similarity between images from their brightness,contrast,and structure.The experiment shows that the improved loss function can provide an improvement in guiding the student network to iterate in the training phase and effectively improve the defect detection effect of the model(3)Based on the algorithm of the thesis,a system for detecting defects on image surfaces is designed and implemented.The system is divided into a client-side and a server-side,with the client-side having the functions of collecting samples,uploading,and displaying defect detection results,etc.and the server-side being able to perform defect detection on samples and serving multiple clients at the same time.