Spatial convolutional self-attention-based transformer module for strawberry disease identification under complex background

• Published in: Computers and electronics in agriculture 2023-09, Vol.212, p.108121, Article 108121
• Main Authors: Li, Gaoqiang, Jiao, Lin, Chen, Peng, Liu, Kang, Wang, Rujing, Dong, Shifeng, Kang, Chenrui
• Format: Article
• Language: English
• Subjects: Image recognition; Multi-head self-attention; Strawberry disease; Transformer module
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2023.108121

•A highly applicable strawberry disease identification method is proposed.•A large-scale strawberry disease dataset has been built.•A spatial convolutional self-attention-based transformer is introduced to extract features. The occurrence of strawberry diseases has a huge impact on the yield and quality of strawberry fruits, resulting in huge economic losses. Real-time and effective identification and diagnosis of strawberry disease is an essential step for strawberry disease prevention. Machine learning-based methods are widely used in strawberry disease identification tasks, but these methods require expertise to design proper strawberry disease feature descriptors. Deep-learning methods have remarkably improved the capability of feature extraction. However, the strawberry disease with complex backgrounds brings great challenges for accurate feature extraction, which leads to poor recognition results of strawberry disease under complex backgrounds. In this paper, an improved transformer-based strawberry disease identification method is proposed to achieve precise and fast recognition of multiple classes of strawberry diseases. First, a multi-classes strawberry disease dataset has been constructed with 5369 images and 12 types of common strawberry disease. To increase the diversity of samples under complex backgrounds, various data augmentation strategies are introduced into the strawberry disease recognition method. Then, Multi-Head Self-Attention (MSA) is used to capture feature dependencies over long distances of strawberry disease images by leveraging the self-attention mechanism. To improve the recognition efficiency, the spatial convolutional self-attention-based transformer (SCSA-Transformer) is proposed to reduce the parameters of the transformer network. The experimental results validated on the constructed strawberry disease dataset demonstrate that the recognition accuracy of the proposed method can achieve 99.10%, which outperforms other methods. Besides, we also observe that the parameters of the classification model are reduced compared with other methods, which effectively improves the recognition efficiency of strawberry diseases.