Research On Rice Image Algorithms Based On Generative Adversarial Networks

In the automation detection of rice quality,the segmentation effectiveness of adhesive rice grains directly affects the subsequent evaluation of grain quality.Traditional machine vision algorithms require feature extraction of rice grains,which overall results in complex calculations and difficulty in ensuring segmentation accuracy.Instance segmentation algorithms based on deep neural network models require the use of template matching mechanisms in designing the network model.The error between the predicted mask and the ground truth mask is computed through a loss function.However,the setting of the loss function often relies on existing segmentation models,leading to the dependence of such models on prior knowledge.Firstly,this paper provides a brief introduction to the importance of nondestructive detection in grain inspection and highlights the need to address the rice image segmentation problem for achieving automated non-destructive detection.Based on the status quo of traditional image segmentation algorithms and deep learning-based image segmentation research both domestically and internationally,the scope and structure of this paper are further clarified.The second part will focus on discussing representative algorithms for solving the segmentation problem of adhesive targets,providing a detailed introduction and analysis of their principles and methods.For image segmentation algorithms based on traditional image processing,clustering-based,watershed-based,and concave pointbased algorithms are introduced.For deep learning algorithms,the principles of Mask R-CNN and YOLACT are discussed.Furthermore,the strengths and weaknesses of current image segmentation algorithms are analyzed,and future research directions are identified.The third part provides a detailed introduction to the core ideas of the YOLO algorithm,with a focus on explaining the principles of the YOLOv5 algorithm and the Generative Adversarial Network(GAN)algorithm.To achieve better rice grain detection without compromising detection speed,a target detection algorithm based on YOLOv5 and GAN is proposed.Compared to YOLOv5,this algorithm requires the addition of only a small number of network parameters during the training phase,but it can bring significant improvements.In the fourth part,the basic principles of the Cascade Mask R-CNN algorithm are introduced.Then,considering the specific characteristics of adhesive rice grains,a generative adversarial training-based segmentation model for adhesive rice grains is proposed,which combines the Swin Transformer as the backbone network for both the generator and discriminator.The model utilizes a mobile window with a multiattention mechanism to extract global features,Cascade Mask R-CNN as the generator model to progressively refine boundaries,Nested-FPN to perform skip connections for multi-scale features,and an adversarial training strategy to guide the generator in learning the mask distribution.Experimental results demonstrate that through adversarial training,the generator can refine boundary masks to a certain extent and achieve high-precision segmentation masks for adhesive rice grains.Compared to other rice image segmentation algorithms,the proposed SWGAN is capable of handling complex scenarios involving rice adhesion and achieving satisfactory segmentation accuracy.Finally,a review of the research content and conclusions of this paper was provided,summarizing the key innovations.