Multi-Category Fruit Image Classification Based on Interactive Segmentation

Abstract

Image classification is the most basic and mature visual task in computer vision. Recently, image classification technology has been widely used. However, a limitation exists in single target recognition and classification tasks for multicategory images. In fruit image classification with complex content of the target image and rich fruit categories, the single use of classification network generation often cannot accurately classify a single-fruit target. To solve this problem, an interactive segmentation-based method for single-category fruit classification in multi-category fruit images is proposed. Herein, an interactive segmentation network and an attention classification network based on deep learning are combined. The interactive segmentation network based on interactive points segments the target to be classified in the image. Then, the classification network identifies and classifies the fruit separately to eliminate the interference of other categories and background information in the image. The classification network is trained on 360 datasets of fruits. The segmentation method before classification can effectively identify single-category fruits in multi-category fruit images. Also, the segmentation and background removal improve the recognition probability of the classification network for a single category of fruit images. Thus, the segmentation method before classification effectively solves single-category fruit classification tasks in multi-category fruit images.