ICAG-NET: A Hybrid CNN–Transformer for Real-Time Plant Disease Identification

Abstract

Plants are among the basic resources available on the earth, as agriculture yields many crops. The diseases of plants can affect global agriculture by influencing the production of agricultural products, and the loss of economic value of crop yield may lead to shortages of food supply. Accurate and quick diagnosis of plant diseases plays a crucial role in avoiding the loss caused by disease and proper treatments. In recent years, CNN-based methods have successfully tackled the problem of plant disease classification. CNNs have an effective approach to capturing the local feature by its convolutions and pooling layers, but the network is unable to capture the long-range dependency. Vision Transformers [14] have been proved to be powerful in capturing the long-range dependencies. However, transformers generally need much higher computational costs, especially for larger-sized images. In this paper, a lightweight fusion architecture of CNN and Transformer, named ICAG-Net, is proposed for real-time plant disease identification. A MobileNetV3-Large [11], which can capture low-level features by convolutions, acts as the backbone network, and a Swin Transformer [23] is employed for modelling long-range contexts hierarchically. Instead of processing the two branches in isolation, an Interactive Cross-Attention Module (ICAM) lets local and global features inform each other during extraction, so each branch benefits from what the other sees. An Attention-Guided Gated Fusion (AGGF) stage then decides, feature by feature, how much weight to give local versus global information, rather than combining the two branches by a fixed rule. To maintain the information in multi-scale feature maps, we introduce a Multi-Scale Feature Integration (MFI) strategy. Experimental results on typical plant disease datasets show that ICAG-Net achieves high accuracy while saving the computational cost, and it holds promise for application in precision agriculture scenarios and edge devices. Keywords: plant disease recognition, hybrid CNN-Transformer, interactive cross-attention, adaptive feature fusion, deep learning, precision agriculture.