Research On Image Segmentation Of Cotton Defects Based On GAN Model

Under the background of great achievements in the process of industrialization in our country.The quality control of industrial products has been paid more and more attention by enterprises and consumers,and there are high requirements for the quality of their products.Especially for cotton products,because of their complex texture,low contrast,weak defects,complex types of defects and other characteristics,resulting in the defect detection process of cotton products becomes particularly difficult.However,the manual detection method is slow,the recognition accuracy is low,and the standard is not easy to be unified.Therefore,the defect automatic detection technology based on computer vision is the inevitable trend of industrial application.This method has the advantages of non-contact,low cost,high precision and fast.In this paper,the accurate segmentation of cotton defects is studied.Several different deep learning models are designed,and the basic methods of image processing are combined.According to the quantitative analysis of the experimental results,the optimal deep learning model is found to realize the rapid detection and accurate segmentation of cotton defects.In this paper,the adversarial idea of generative adversarial networks is adopted.The existing semantic segmentation network is improved and optimized,and two different ideas are adopted to design and train the model.Firstly,we use adversarial training for semantic segmentation networks,and use FCN,U-Net and Seg Net as generators.By training the adversarial semantic segmentation networks,the segmentation image is generated directly from the defect image.The other method is to achieve defect segmentation based on image repair mechanism.A lightweight and efficient defect repair model is designed,and conduct adversarial training on the model.In this model,the defective image is restored to a non-defective image,and then the image processing algorithm is used to obtain the segmentation image of the defective area.The research content of this paper mainly includes the following three parts.(1)Sample preprocessing and enlargement methods.In view of the small number of effective defect samples,the existing cotton samples are screened and cut.Then the image processing operations of rotation,flip,transpose and random synthesis are used to enlarge the quantity of cotton samples,so that a large number of defect samples can be obtained quickly without heavy manual labelling.Finally,a sample set of effective cotton is obtained.(2)Semantic segmentation method has been studied for defect segmentation.Firstly,different semantic segmentation networks are studied,including their network structure,detailed characteristics and advantages and disadvantages of models.This paper studies three semantic segmentation networks,FCN,U-Net and Seg Net,and then the training process of the combination of segmentation network and generative adversarial networks are researched and designed.The optimal segmentation result is found by adjusting the super parameters of the training model.Compared with the segmentation results of general segmentation network,the advantages of the adversarial semantic segmentation networks are analyzed through the experimental results.(3)A defect segmentation method based on image repair is studied.A defect repair model is designed with the structure of encoder and decoder to realize the defect image repair.The repair model is lightweight and can achieve good repair effect.Finally,the repair image and the defect image are compared and analyzed,and the image processing method is used to complete the accurate segmentation of defects.The results show that the performance of the proposed adversarial semantic segmentation networks is improved for the general semantic segmentation networks.Seg Net has the best segmentation accuracy after using adversarial training,and its Dice score reaches 0.8195.The model size of the adversarial repair network in this paper is only 14.4MB,and the Dice coefficient reaches 0.8696,which has a higher segmentation accuracy than other segmentation networks in this paper.The model not only ensures light weight,but also has higher real-time performance,and it has good application value.