Category Detection Of Strip Surface Defects Based On GAN And Lightweight Neural Network

Strip steel is an important industrial raw material,and strip steel defect detection is critical to product quality.Deep neural network based on big data training can achieve accurate detection,but the amount of data collected under the restriction of various factors in industrial environment often cannot meet the training needs.GAN is an important data set expansion method,but it is difficult to train on generating high-quality multi-category images.In addition,in order to perform real-time detection in the industrial field,the neural network must be as light and fast as possible.Therefore,this paper focuses on GAN and lightweight neural network to research the category detection of strip steel surface defects.In order to make GAN generate images with clear categories and good quality to expand the strip steel defect dataset,firstly,four generators are designed that can strengthen the guiding role of image labels in different ways.Secondly,a discriminator is designed which can not only judge the authenticity of images but also judge the categories of images.Then,the objective function is designed that can enhance the training stability and alleviate the mode collapse problem.Finally,the generator,discriminator and objective function are combined to form four GAN models,and the best GAN model is selected through experiments to expand the dataset.In order to meet the accuracy and real-time requirements of industrial detection,Firstly,the principles of four representative lightweight image classification networks are studied,including Shuffle Net V2,Mnas Net,Mobile Net V3 and Ghost Net.Then,according to the actual situation,these four lightweight models are further streamlined and optimized.Finally,the image classification models are trained on the expanded dataset and the model with good comprehensive performance is selected for the category detection of strip steel surface defects.The experimental results show that all the GAN models designed in this paper can generate images with clear categories and good quality.Among them,the average FID of the CDGAN(Conditional Deconvolution Generative Adversarial Network)used to expand the strip steel surface defect image dataset is 6.22.All the lightweight neural networks improved in this paper can achieve image classification well.The selected improved Ghost Net can effectively realize the category detection of strip steel surface defects.Its accuracy is 95.33%,its parameter is 0.09 M and its inference time is 23.1 ms.