Research On Vision Inspection Technologies Of Product Surface Defect Detection Based On Deep Learning

In recent few years,great progess has been seen in deep learning theory.More and more high-precision models were proposed,achieving great success in many tasks.However,it is still extremely hard for deep learning to be deployed in industrial scenarios,rare deep learning algorithms have been used in large-scale application.There are two main reasons.First,deep models often require massive labeled examples to avoid over-fitting problems.In actual industrial production,the number of examples is limited,and the cost of obtaining examples is high.Second,deep learning consumes a lot of computing resources and requires highperformance computing equipment.In industrial environment,model usually needs to be deployed on resources-constrained equipment such as embedded equipment and industrial personal computers,and is very sensitive to cost.For the application problems of deep learning in industrial production,this article takes the detection of button surface defects as an example and proposes a complete solutions.First,data is preprocessed to reduce the redundant information and improve the quality of the data set,then overfitting and computing performance problems are solved by transfer learning and model pruning,and finally an online detection system is designed and proposed.The experiment proves that the method and system in this paper can be applied in the production environment and achieve good results.The main research contents and results of this article are as follows:(1)Processing methods of industrial image,including data preprocessing and data enhancement.Data preprocessing uses traditional digital image processing methods to perform background subtraction,filtering,extraction of regions of interest,cropping and stitching on the images collected on industrial environment,reduce redundant information in the image.Method also clean and segment the dataset to improve the dataset quality.Data enhancement uses a variety of traditional digital image processing methods and Mixup to enhance the industrial datasets,which can effectively reduce the risk of overfitting and improve the generalization ability of the model.(2)Transfer learning methods.In view of the limited number of examples and training difficulties,a method of transfering pre-trained models for defect detection is adopted.Model structure design and hyperparameter selection in the transfering process are studied,and on this basis,according to the characteristics of industrial datasets,a transfering and fine-tuning method based on auxiliary tasks is proposed.Experiment shows that the accury and recall rate of proposed method can reach about 95%,which satisfies the requirements of industrial production.(3)Model compression method.In view of the large amount of redundancy in the model after transfering,a channel pruning method is proposed.This method uses feature map activation as the channel selection criterion within the layer,and loss estimation as the channel selection criterion between layers.Compared with the traditional method,this method can realize the adaptive pruning between channels while controlling the overall pruning ratio,without having to set the pruning ratio of each layer with experience,which is more flexible.Experiments show that this method can compress floating point operations and parameters of the transfered model by 3.2x and 8.3x respectively,without affecting the generalization performance of the model.For the VGG16 trained on Cifar-10,this method can also compress the floating point operations and parameters by 2.5x and 5.85 x without significantly reducing the accuracy(-0.42%).(4)Design of detection system.Based on above methods and experiments,a complete hardware and software system are also designed and implemented.The hardware is a sorting device,which can automaticly load material,capture images and sort material.The software system uses MVC architecture and producer-consumer model with multi-thread and multiprocess.The system satisfies the requirements of industrial production for accuracy,high speed and reliability,and can be practically applied in industrial production environments.