Lightweight Vision Transformer Compression for On-Device Plant Disease Detection in Resource-Constrained Agricultural Field Conditions
A research paper proposes a unified Vision Transformer compression framework combining Hessian-Balanced Adaptive Block Pruning (H-BAC), quantization, and attention-based knowledge distillation for on-device plant disease detection. The study evaluates each technique independently through controlled ablation studies and integrates the best-performing components into a sequential deployment pipeline for agricultural constraints. The work targets chilli (Capsicum annuum) disease classification using a 3-class village-split dataset.
The paper addresses the challenge of deploying high-accuracy Vision Transformers on resource-constrained devices for chilli plant disease detection in India. It introduces a compression framework that combines Hessian-Balanced Adaptive Block Pruning, quantization, and attention-based knowledge distillation. Each technique is first evaluated independently, then the best components are integrated into a sequential pipeline tailored to real-world agricultural field conditions.
The framework uses second-order sensitivity estimation via Hessian-Balanced Adaptive Block Pruning to guide pruning decisions, combined with quantization and attention-based knowledge distillation. The sequential integration of independently optimized compression techniques suggests a modular approach to balancing accuracy and computational efficiency for edge deployment.
This research targets agricultural applications where expert disease identification is scarce and on-device inference is necessary due to field connectivity constraints. The focus on a village-split dataset indicates an emphasis on generalization across different geographic or demographic conditions, which is critical for real-world deployment in developing regions.
If successful, the compressed ViT could enable low-cost, offline plant disease detection tools for farmers, reducing reliance on expert consultation and potentially improving crop yields. The approach may also be applicable to other edge AI tasks in agriculture or similar resource-constrained domains.
Next observable signals include publication of detailed ablation results, benchmark comparisons against existing compression methods, and potential open-source release of the compressed model or code. Further validation on additional crops or disease classes would strengthen the framework's generalizability.