Resmobnet: A Lightweight Dual-Branch Deep Learning Architecture with SE Attention for Cotton Plant Disease Identification
Resmobnet: A Lightweight Dual-Branch Deep Learning Architecture with SE Attention for Cotton Plant Disease Identification
Digvijay Singh *1, Sushil Kumar Sharma*2
*¹Research Scholar, Department of Computer Science and Engineering, Institute of Technology and Management Aligarh University: AKTU Lucknow. Email: dsinghrana876@gmail.com
*² Assistant Professor, Department of Computer Science and Engineering, Institute of Technology and Management Aligarh University: AKTU Lucknow. Email: sushmca@gmail.com
Abstract— Cotton is a staple food crop in many developing nations but yield is often decimated (20-40%) by diseases due to late or wrong diagnosis in the field. In this work, we present ResMobNet, a bespoke hybrid convolutional neural network architecture that combines MobileNetV2 depthwise-separable architecture efficiency, ResNet-style residual connections and Squeeze-and-Excitation (SE) channel attention mechanisms within a multi-scale feature fusion approach using a dual-path structure. Using the publicly available Kaggle Cotton Disease Dataset (1951 images, four classes), the proposed model attains 98.31% test accuracy, 98.33% weighted precision, 98.31% recall, 98.30% F1-score and a macro-averaged AUC of 0.9987, outperforming ResNet50, regular MobileNetV2 and VGG16 models by 2.03, 2.70 and 4.05 percentage points respectively. ResMobNet offers a good accuracy-efficiency tradeoff with just 6.58 million parameters and 42.3 ms inference time on the CPU. A rigorous ablation study verifies isolated impacts of 1.17 pp due to SE attention, 1.35 pp due to residual blocks, 2.03 pp due to the double-branch design and 2.57 pp due to the two-stage transfer learning protocol. Grad-CAM results justify biologically meaningful localisation of diseased areas. The model is exported in quantised TFLite format (6.70 MB), for direct use on mobile edge devices for real-time precision agriculture.Keywords— Cotton disease detection; deep learning; MobileNetV2; ResNet; squeeze-and-excitation attention; multi-scale feature fusion; transfer learning; Grad-CAM; precision agriculture; TFLite..