Research On Intelligent Detection Of Rice Appearance Quality Based On Machine Vision

One of the main ration types in China is rice,and the price is directly correlated with its processing accuracy and broken rice content as quality indicators.In addition,the research and application of moderate processing technology is a crucial element to strengthen the savings and loss reduction across the whole food industry chain.This is essential to ensuring food security.In this paper,the processing accuracy of rice,broken rice,and yellow grain rice were studied and the dataset was built by selecting standard samples specified in national standards,and the detection approaches defined in the national standards were used for comparison.Further to the investigation of image pre-processing and adhesive rice grain segmentation techniques,two classification models were built.They are a machine learning-based rice appearance quality identification model and a deep learning-based classification model.Simultaneously,a machine vision-based rice appearance quality index evaluation method was developed to achieve intelligent detection of rice appearance quality.These techniques can be applied to the field of flexible rice processing to guide enterprises to moderate processing and improve rice supply quality.The following is the focus of this essay:(1)The fact that the current image capture technique differs from the real detecting process of rice grains adhering and scattering randomly.In light of the aforementioned circumstance,this paper employs a vertical separator to lessen rice grain stickiness as well as a flatbed scanner to obtain multi-grain mixed attitude rice RGB photographs.This work analyzed and researched the usual pre-processing and adhesion rice segmentation techniques to remove noise and some adhesion issues from the recorded images.Finally,efficient algorithms for the H-component grayscale method,median filtering,and adaptive threshold segmentation which are most suitable for the characteristics of the research object are presented.Additionally,a convex hull defect detection-based approach of two rice grain adhesion segmentation was proposed to realize the segmentation of rice single-grain images.The segmentation accuracy reaches 94.99% when the advantages of the equipment and the proposed algorithm are combined and the result can satisfy the needs that follow.Meanwhile,the image data is enhanced to provide enough data samples for subsequent deep learning models.(2)There are issues with different inspection links having different focuses,and the same session of index detection needs to be tested separately when it comes to rice appearance quality inspection.In this study,a thorough approach to rice appearance quality identification based on multi-class fusion(color,texture,shape)feature extraction was provided as a solution to the aforementioned issues.And an 8-21-5 structured Back Propagation Neural Network(BPNN)model was created to accomplish the comprehensive recognition of rice appearance quality.The model’s recognition accuracy reached 86% for five different categories of rice.The findings demonstrate that the detection approach based on conventional feature extraction and BPNN is feasible for the comprehensive recognition of five different types of rice.However,the traditional manual design and feature extraction process is complicated and onesided,which leads to its poor recognition and classification effect.(3)The IR＿HDC＿NAM model was bulit in this work given the complexity of the traditional manual feature design process and the lopsidedness of manually extracted features leads to poor classification results.The model introduced the Inception-v3 structure,hybrid dilated convolution,and Normalization-based Attention Module over the Res Net34 model,and was able to automatically extract the rich detailed features of rice to improve its classification effect.In addition,it is compared and analyzed with the classical Convolutional Neural Network(CNN)model.According to the experiments,the suggested model outperformed Alex Net by 3.2% and Res Net34 by 1.2 percentage points,achieving a classification accuracy of 97.41%.The model took 28 s for each group of 5800 images,and the outcomes can satisfy the demands for quick,non-destructive,and precise testing of rice appearance quality.(4)In the current study of broken rice rate and yellow grain rice content,using the individual ratio rather than the mass ratio in the national standard results in larger errors in the findings.According to the above identification model,the relational expressions of the projected area and mass of rice were discovered in this survey by designing a relational model to solve the problem.Then the calculation rules in the national standard were combined to realize the calculation of processing accuracy,broken rice rate,and yellow grain rice content for the whole batch of rice.There is a relative error of less than 1.3% between the constructed machine vision-based detection method of broken rice rate and yellow grain rice content and manual detection.Moreover,compared to the national standard approach,the method described in this work has little bias in the detection of quality indicators.It can contribute significantly to the methodical and objective assessment of rice surface quality.