The internal and external quality of green plum has an important impact on its processing process. Conventional manual sorting not only has low classification efficiency, but also is difficult to realize standardized operation due to personal subjective factors, which can not meet the market requirements. In the aspect of defect classification, based on deep learning technology the vision transformer network was used in machine vision system, which introduced multihead self-attention to improve the global feature representation ability, and reduce the gradient through the softmax function to realize the detection and sorting of multiple categories (rot, crack, scar, spot and normal) on the surface of green plum. The results showed that the discrimination accuracy of rot, scar, crack and normal plum images reached 100%, spot reached 97.38%, the average discrimination accuracy was 99.16%, and the average test time of each group was 100.59ms. The discrimination accuracy and average discrimination accuracy of this network were significantly better than VGG and ResNet-18 network. In terms of internal quality (SSC) prediction of green plum, based on hyperspectral imaging technology, the LRTR-SCAE-PLSR prediction model of green plum was constructed by combining the denoising advantages of LRTR and the dimensionality reduction advantages of SCAE. The results showed that when the network scale was 119-90-55-36, RP was 0.9654 and RMSEP was 0.5827%. By comparing the two dimensionality reduction models of SCAE and LRTR-SCAE, LRTR-SCAE model not only had lower dimensions, but also significantly improved the correlation coefficient of prediction set, which verified the dimensionality reduction and denoising advantages of LRTR-SCAE model. An intelligent equipment for nondestructive sorting of internal and external quality of green plum was designed and built. The whole machine had small size and simple structure. The sorting results met the requirements of green plum deep processing.