Are Transformers more Suitable for Plant Disease Identification than Convolutional Neural Networks?

Abstract

Recent studies suggest that the transformer-based architectures have made unprecedented achievements in various computer vision tasks, such as image classification, object detection, semantic segmentation, etc. However, as the de facto approach, convolutional neural networks (CNNs) have reigned for a decade in plant disease identification tasks. We cannot help but propose a scenario: are transformers more suitable for plant disease identification than CNNs? Conceivably, this work aims to further investigate and evaluate whether it is feasible to trivially switch to transformers in plant disease identification tasks. To address this issue, a series of experiments were curated into three training strategies and implemented on two representative plant disease image datasets. Experimental results indicated that the vanilla transformer methods performed unsatisfactorily when trained from scratch, and the two mainstream transfer learning-based approaches achieved better prediction performance but with little margin. It is encouraging that the self-supervision-based transformer methods outperform their CNN-based counterparts, which provides new insights for agricultural image processing. Importantly, this work revealed a model selection solution based on the experimental results and target datasets, and gained a better interpretation by visualizing the feature maps.