Multi-organ Jujube Classification Based on a Visual Attention Mechanism

Abstract

Jujube variety classification is a challenging task because of the difficulty in identifying discriminant features, making it difficult to find the subtle features that can fully represent the variety. Besides, jujube identification using single-organ fruit is not sufficiently reliable because different jujube varieties usually have similar fruit shape. To overcome these problems, this paper proposed an automatic jujube identification model based on attention mechanism by combining multiple organs of jujube. The model used a conventional neural network to perform feature extraction on images, and subsequently adopted some fusion techniques to further process the feature maps. By introducing the attention mechanism, the model could recalibrate channel and spatial characteristic responses adaptively so as to focus on the more discriminative regions of the images. Based on the idea of fusing multi-organ features, the network effectively obtained more significant cues for jujube recognition. Experimental results showed that the proposed network had a higher accuracy of 94.77% on jujube classification compared with other methods. It is demonstrated that the network was of great value to jujube recognition research.