Efficient Triplet Attention Network for Fine-Grained Crop Disease Classification

Abstract

Fine-grained crop disease classification is a challenging task due to problems such as visual interference and scene complexity. In recent years, attention mechanisms have received increasing attention for their superior function in improving the performance of deep CNNs. This paper presents an efficient triplet attention (ETA) module, which captures channel attention and spatial attention information more effectively to improve the description ability of deep CNNs. In addition, high-quality data samples can also bring benefits to network training, and CutMix is a common method to obtain such samples. However, CutMix suffers from the problem of label misassignment due to random cuts. Inspiration from this, we extract attention information from the output features and propose the AttentionMix data augmentation strategy, which can effectively solve the label misallocation problem in CutMix. The ETA module and AttentionMix can work together on deep CNNs to obtain larger performance gains. We validate the effectiveness of our method on common plant pest and disease classification datasets and our own crop disease datasets. Experimental results of the proposed method convincingly show the advancement compared to the state-of-the-art methods.

Keywords: Fine-grained crop disease classification, CNNs, Channel attention, Spatial attention, Data Augmentation