Research on recognition of stem/calyx and defects of dried Hami jujube based on ResNeXt and transfer learning

Abstract

Objective: This study focuses on solving the problems that stem and calyx are mistakenly recognized as defective jujube in the current detection process of jujube. Methods: A method for identifying stem / calyx and defects of dried Hami jujube based on deep learning and image processing was proposed. By improving the depth residual network ResNeXt-50, using the region of interest extraction method and transfer learning technology, TL-ROI-X-ResNeXt-50 classification model was proposed to realize the classification of stem / calyx and defects of dried Hami jujube. Results: Through the comparison of model experiments, the region of interest extraction method and transfer learning technology could reduce the calculation cost of the model and improve the accuracy. The accuracy of model recognition could reached 94.17%. Conclusion: The method can initially meet the production requirements of on-line detection equipment for stem / calyx and defects of dried Hami jujube, and provide theoretical basis and technical reference for the development of rapid nondestructive detection system for stem / calyx and defects of other similar fruits.

Keywords

• hami jujube
• image identification
• deep residual network
• transfer learning
• interest extraction