AgroVisNet and BD-PlantDX: Compact Vision Architecture and Expert Dataset for Edge Crop Disease Diagnosis
Original title:AgroVisNet: A lightweight Convolutional Network and the BD-PlantDX Expert-Validated Benchmark for Radish, Potato and Pointed Gourd Disease Classification
To address the lack of domain-validated local data and resource constraints in rural agriculture, researchers introduced BD-PlantDX—a 12,432-image benchmark covering radish, potato, and pointed gourd diseases in Bangladesh—alongside AgroVisNet, a compact convolutional architecture. With just 290,572 trainable parameters, AgroVisNet combines grouped bottleneck residual blocks, spatial-channel attention, and multi-scale depthwise convolutions. When quantized to full integer precision, the model shrinks to 0.46 MB while preserving a 99.52% test accuracy and executing in 8.40 ms on a single CPU, bringing botanical diagnostic capabilities directly onto inexpensive offline hardware.
Why it's worth reading
It demonstrates practical edge vision by compressing an expert-validated crop diagnostic model to 0.46 MB, offering a viable blueprint for offline AI deployment on ultra-low-cost devices.