Visual classification of apple bud-types via attention-guided data enrichment network

Abstract

The number of flower buds on the apple tree is the crucial factor for fruit load determining, thus the essence of apple tree pruning is bud removal. Most horticulture activities in apple orchards at present primarily rely on skilled farmers. However, distinguishing between different types of apple buds is still hard work for many planters due to their similar appearances. The most recent published works have proven the superiority of computer vision and deep learning in image recognition tasks. Deep convolutional neural network (DCNN) is an efficient type of network in deep learning architecture for visual features analysis. To categorize types of apple bud at the fine-grained level, a DCNN-based visual classification model denoting the attention-guided data enrichment network (ADEN) is proposed. Specifically, in ADEN, the ResNeSt50 network is used as the feature extractor module for characterizing the apple bud trait from each input image. Based on attention maps, the attention-guide data enrichment module, containing attention-guided CutMix and attention-guided erasing, is designed for the task of enriching training samples via dropout and fusing local features of images, which further improves the training efficiency and discriminative ability of the classifier. All the experiments are conducted on the orchard-shot image dataset contained two classes of apple buds, include the flower bud and the leaf bud. The proposed method conveys a consistent and significant improvement in performance and achieves testing accuracy of 92.39% with satisfying precision, recall and f1-score, which outperformed the comparative models. The proposed method can readily realize accurate identification for bud-types of apples and is helping to promote the advancement of pruning and training robotization in orchards.