Research On Non-destructive Detection Algorithm Of Soybean Damage Based On Computer Vision

As one of the major high-protein food materials and oil crops in China,the superior and stable yield of soybean is of great significance.Seeds are the basis for soybean breeding and production,and their quality directly affects the final yield and output quality of the crop.Existing seed screening methods suffer from low accuracy,slow speed,and tend to cause secondary damage to the seed.This dissertation combines computer vision technology with the characteristics of fast,nondestructive and high throughput,integrates the use of near-infrared spectral imaging and visible light imaging technology to study the nondestructive detection algorithm of soybean damage,and establishes an accurate,fast and nondestructive soybean detection model.The main problems and challenges of nondestructive detection of soybean are studied in depth from three aspects:single grain detection,low throughput detection,and high throughput detection.The contents and innovations of this dissertation are as follows:1.Research on machine learning-based algorithms for soybean near-infrared spectral data identification.Near-infrared spectroscopy is mostly used for crop seeds such as corn,rice and wheat,and there is little research on soybean seeds with poor light transmission.In this dissertation,the FT-NIR spectrometer is used for non-contact data acquisition.Aiming at the problems of large amount of information,invalid information and redundant information in near-infrared spectral data,the first derivative combined with SG smoothing is proposed to preprocess the original spectral data,so as to select the effective band.To address the problems of weak interpretation of physical layer overlap information in spectral data and the high dependence of the model on mathematical models,we propose to use machine learning methods for feature extraction and classification,and achieve accurate detection of soybean seed coat cracks using random forest and variable selection optimization algorithms.The accuracy of the random forest model proposed in our dissertation is 80%,and the detection accuracy of the original model combined with the variable selection optimization algorithm is 84%.This study demonstrates the effectiveness of using machine learning combined with NIR spectroscopy for nondestructive detection of soybean seed coat cracks.2.Research on low throughput soybean detection algorithms based on multi-model fusion.The use of machine learning combined with NIR spectroscopy has the following shortcomings:the equipment can only detect single soybean seeds,but not multiple seeds simultaneously;the NIR spectral data can produce widely varying results under different pre-processing methods,which affects the accuracy of the classifier to a certain extent;machine learning methods have poor robustness and a high number of model hyper-parameters.To address the above problems,this dissertation designs a two-stage detection strategy with multi-model fusion,i.e.,a segmentation-classification strategy for soybean detection.The first stage uses Mask R-CNN model for accurate detection and segmentation of soybeans with low throughput.The classification accuracy in this stage is poor because the soybean damage forms are highly similar.To address the problem of poor classification accuracy,the second stage uses a lightweight model,SNet,which utilizes a hybrid feature recalibration module to improve the classification performance of the model with a parametric number of only 1.29 M.The experimental results show that SNet achieves accurate classification of soybean seeds with 96.2%accuracy,and its performance is better than other comparative methods.3.Research on high-throughput soybean detection algorithms based on invertible convolutional networks.High-throughput detection suffers from the following problems:high time and labor costs for fine labeling of dense target data;large number of soybean seeds with a high probability of adhesion and occlusion between seeds;and information loss in the classical feature pyramid network in the detection model.To address the above problems,firstly,a self-supervised approach is designed to construct a large amount of synthetic data and automatically generate pixel-by-pixel annotation information,which solves the complex problem of fine annotation while alleviating the problem of scarce agricultural data and annotation information to a certain extent.Secondly,the Inv-Mask model based on invertible convolution is proposed to achieve accurate and fast detection of high-throughput seeds by using the features of invertible convolutional bidirectional mapping and low complexity of spatial and channel domains.Finally,a feature selection pyramid network(FS-FPN)based on invertible convolution is designed,which enhances the ability of extracting multi-scale features from the model and solves the problem of model accuracy degradation caused by information loss.The experimental results show that the synthetic data method proposed in this dissertation effectively solves the problem of data shortage,and the designed Inv-Mask model combined with FS-FPN has good performance,and its detection accuracy and detection speed are significantly improved,which has a greater practical value.