| Green plums are not suitable for direct consumption,and most of them are eaten after processing.Different deep-processing products of green plum have different requirements for the composition characteristics of raw fruit.During the deep processing of green plum,the raw fruit classification is mainly based on the defects of green plum,sugar and acidity,etc.The conventional manual sorting is not only inefficient in grading,but also difficult to achieve standardized operation due to personal subjective factors,which cannot meet the market demand.And the traditional physical and chemical detection methods are destructive and inefficient,which cannot meet the actual detection requirements.In order to solve the problems of low manual sorting accuracy,inconsistent standards and destructive detection of green plum components,the appearance defects,soluble solid content(SSC)and hydrogen ion concentration(p H)of green plum are taken as the research objects in this paper.Based on artificial intelligence technologies such as multi-spectral imaging,machine vision,3D reconstruction,deep learning,etc.,and various technical advantages are integrated to obtain accurate image information and spectral information of green plum.The classification model of appearance defects of green plum based on deep learning,the internal quality prediction model of green plum based on multi-channel band group,and the three-dimensional model of green plum based on improved SFM are established.Green plum quality intelligent sorting equipment are built to achieve high-precision and low-cost green plum appearance defect and internal quality sorting.The green plum with high sugar content and normal surface is selected for fresh food,the green plum with high acidity and low sugar content is used for making plum essence,dried plum,crisp plum,etc.,and the green plum with low acidity and high sugar content is used for making plum juice drinks,fermented plum wine,etc.The research on intelligent separation technology and equipment is of great significance to improve the separation efficiency,ensure the separation quality,and increase the added value of green plum products.The specific research contents of this paper are as follows:1.In order to solve the problems of high hardware cost and low sorting efficiency of hyperspectral nondestructive testing technology for green plum,the study on the prediction of sugar and acidity of the internal quality of green plum based on multispectral technology was carried out.SSC and p H were taken as the research object,the VIS-NIR and SW-NIR full-band spectral information of green plum were collected,the spectral data was corrected and pre-process.The improved random forest algorithm based on inductive random selection(IRS-RF)was proposed to screen four groups of characteristic bands.The BO-Cat Boost model based on Bayesian optimization algorithm was constructed to predict the SSC and p H content of different green plum.The experimental results showed that the dimension of characteristic bands screened by MSC+IRS-RF was the lowest,and the lowest of VIS-NIR and SW-NIR bands was 53 and 100 in SSC prediction,and the lowest of VIS-NIR and SW-NIR bands was 38 in p H prediction.MSC+IRS-RF+BO-Cat Boost model was superior to PLSR,XGBoost and Cat Boost regression models in SSC and p H prediction,R_P~2 were up to 0.957 and 0.982.Based on the above research contents,a multi-spectral collection platform was built.The multi-channel spectral sensors were selected to obtain the spectral information of green plum,and the spectral data was preprocessed.BO-Cat Boost was used to establish the prediction model of the internal quality(SSC,p H)of green plum based on the multi-channel spectral sensors.The R_P~2 about SSC and p H prediction of green plum based on the spectral data of 48 characteristic wavelengths screened by four multi-channel spectral sensors was 0.945 and 0.973,which met the actual sorting requirements.2.In order to solve the problems such as large amount of calculation and slow convergence of the model in the classification of complex and diverse appearance defects of green plum,the study on the classification of appearance defects of green plum based on deep learning technology was carried out.According to different processing requirements of green plum,the appearance defects of green plum were reasonably classified,and the green plums were divided into five categories:rot,scar,spot,crack and normal.The data set production and image preprocessing of appearance defects were introduced.The results and analysis of the appearance defect classification of green plum based on Vi T network were carried out.The Vi T network for the defect classification of green plum has good performance in various defect classification,and the classification effect was significantly better than the VGGNet network and Res Net-18 network.But the running time became longer due to the increase of parameter quantity.In order to solve the above problems,the classification results and analysis of plum appearance defect of different Swin Transformer deep learning networks were carried out.By introducing local attention and reducing the computational complexity of the model,the classification performance of Swin-B network was the best.Based on the above research,the Layer Scale-Swin B-RAdagrad network was proposed.The experimental results showed that it has a good comprehensive performance in the classification of five green plum appearance defects.Compared with conventional Swin B,the average classification accuracy was improved by 2.0%,and the average test time was shortened by 44.2%,and the accuracy of defect classification of scar,rot,crack and spot have been improved by 8.5%,0.7%,4.4%and 1.9%.3.In order to solve the erroneous and missed judgment of surface micro-defects detection of small fruit such as green plum,and further improve the accuracy of appearance quality classification of green plum,which caused by two-dimensional image acquisition angle,light change,lens distortion,etc..The platform of green plum image acquisition device was built to obtain the two-dimensional image sequence of green plum from multiple perspectives.The improved SFM algorithm were proposed to accomplish 3D reconstruction of green plum,which obtained high reconstruction accuracy and met the requirements of green plum micro-defect(spot)detection.For the dense point cloud model of green plum,an improved adaptive segmentation algorithm based on Lab space was further proposed to achieve effective segmentation of micro-defect(spot)point cloud.The experimental results showed that the average running time of the improved adaptive segmentation algorithm was 2.56s,while the average running time of the traditional algorithm was 6.91s,which was reduced by 63%.The running time of the algorithm was greatly reduced and good segmentation results was obtained.Through the Euclidean clustering of spot cloud,the defect information of green plum spot was extracted based on the consistency of random sampling.The experimental results showed that when the density of sampling points was80,the relative error of curve contour area was only 0.0066.4.According to the sorting requirements of green plum,the hardware system of intelligent sorting equipment was mainly composed of appearance defect classification module,internal quality sorting module,conveyor,equipment base,sorting mechanism,drive motor,control system,etc.The software system was mainly composed of user management,image data preprocessing,image acquisition,internal quality sorting,appearance defect classification,system setting and real-time operation.The hardware composition and work flow of each module were introduced,the sorting standard was introduced,and the man-machine interface for quality sorting was built.The sorting results of green plum was displayed by the system in real time to realize the visualization of detection results.The performance indicators of the intelligent sorting equipment for green plum quality met the requirements of the project by system testing,and the feasibility of real-time detection of the intelligent sorting equipment for green plum quality was verified. |
PDF Full Text Request