Classification of jujube defects in small data sets based on transfer learning

Abstract Although convolutional neural networks have achieved success in the field of image classification, there are still challenges agricultural product quality sorting such as machine vision-based jujube defects detection. The performance defect detection mainly depends on feature extraction and classifier used. Due to diversity materials variability testing environment, traditional method manually extracting features often fails meet requirements practical application. In this paper, a model small data sets based network transfer learning is proposed actual demand Firstly, original images collected from production line were pre-processed, augmented establish set five categories defects. CNN then improved by embedding SE module using triplet loss function center replace softmax function. Finally, depth pre-training ImageNet was used conduct training set, so that parameters could fit parameter distribution image, transferred complete realize classification results visualized heatmap through analysis accuracy confusion matrix compared with comparison models. experimental show SE-ResNet50-CL optimizes fine-grained problem recognition, test reaches 94.15%. has good stability high recognition complex environments.