A study on recognizing apple surface scars using residual network and transfer learning

Abstract

In this study, a method for apple surface scar recognition based on residual networks and transfer learning is proposed. The dataset was constructed by collecting healthy and disease-infected apple and leaf images and expanded using data enhancement techniques. Using ResNet34 as the base model, multiple improved network models are obtained by combining migration learning, and finally the accuracy of apple surface scar classification is successfully improved to 99.7%. The experimental results show that migration learning improves the model accuracy in small samples, while data augmentation also significantly improves the model performance. The model is able to accurately detect scabbed fruits and healthy fruits, which can promote the development of China’s fruit grading industry and provide a useful practical basis for agricultural science and technology and rural revitalization.