Lightweight Image Classification Model

Identify Grain Defects

Achieves high recognition accuracy while significantly reducing model complexity and computational cost, enabling real-time non-destructive inspection on resource-constrained devices.

Yunzhao, Ma · Wu, Wenfu · Yan, Xu

Frontiers in Plant Science 2026

Specifications

Recognition accuracy

{'zh': '94.2', 'en': '94.2'} %

Parameter reduction

{'zh': '30.3', 'en': '30.3'} %

Computational cost reduction

{'zh': '27.6', 'en': '27.6'} %

Advantages

Accuracy improves

Accuracy increases by 6.1% to 94.2%, showing that lightweighting does not sacrifice recognition capability.

Model is smaller

Parameter count drops by 30.3%, meaning lower storage and memory footprint and easier deployment.

Runs faster

Inference speed rises from 155 FPS to 183 FPS, keeping pace with real-time inspection.

Applications

• wheat grain quality inspection：Can be deployed on inspection lines to automatically sort six grain categories, replacing manual visual inspection.
• edge computing devices：Large reductions in parameters and computation allow real-time operation on edge devices without cloud offloading.
• smart agricultural devices：The lightweight model fits into smart agricultural machinery or sorting equipment for on-site quality checks.
• real-time visual inspection systems：At 183 FPS, the model meets the frame-rate demands of real-time visual inspection and integrates easily with existing systems.