Apple appearance quality classification method based on double branch feature fusion network

Abstract

This paper carried out a classification study on the quality of different types of apples. In order to achieve a high-precision classification of high-quality apples in various dimensions, shorten the classification time, improve efficiency, and make it feasible for practical applications. This paper proposes an automatic recognition and classification model OB-Net (for apple appearance quality) based on a dual-branch structure. The model consists of two branches, an O branch and a B branch. The O branch extracts the shape and contour features of the apple surface defect by decomposing the feature map into high and low frequencies. The B branch extracts the shape size and texture characteristics of the apple surface defect by fusing channel attention and spatial attention mechanisms; this is done to further increase the feature distance between different types of defects. Experimental results show that the classification accuracy rates of rot, insect bites, russeting, scratches, and intact apples are 95.65%, 98.17%, 94.62%, 92.02% and 97.67%, respectively; the overall classification and recognition accuracy rate reaches 95.64%. Finally, the three aspects of the feature map, heat map and category probability statistics map extracted from the OB-Net network model prove the validity of the network model in the classification of apple appearance quality.