FSLNet: Filter sensitivity-based lightweight network for rice leaf disease recognition

Abstract

Leaf disease recognition is critical for guaranteeing rice quality and yield. However, hindered by similar symptoms and complex background interference in practical field scenarios, existing models face the significant challenge of balancing accuracy and lightweight requirements for edge devices. To address the challenge, this paper proposes a filter sensitivity-based lightweight network (FSLNet), which comprises three key modules: a filter sensitivity evaluation algorithm (FSEval), a sensitive channel spatial attention mechanism (SCSAM), and a sensitivity-driven model compression method (SDMC). To overcome similar symptoms, FSEval is designed to calculate each filter’s sensitivity to different diseases. Then, the channel penalty is imposed on non-sensitive filters (CP-NSF) to make FSLNet extract differential features. To mitigate background interference, during each training epoch, SCSAM can adaptively infer an attention map only from sensitive channels containing differential features, to optimize their weight distribution. To accommodate edge devices in the practical field, SDMC incorporates a sensitivity-driven channel penalty pruning (SDCPP) strategy and a fine-tuning method with a low pruning teacher model via knowledge distillation (FT-LPT-KD) to build a lightweight model. Experimental results on the self-built dataset and public Paddy Doctor dataset demonstrate that FSLNet has only 0.38M parameters, just 1/6th of the smallest benchmark GLDCNet, while can achieve an average accuracy of 95.26% and 97.98%, respectively, which is 1.84% and 0.77% higher than those of the state-of-the-art schemes. Further testing on a resource-limited edge device reveals that FSLNet not only outperforms existing methods in recognition accuracy but also can achieve a real-time recognition speed of 28 FPS.

Plant phenotyping relevance

イネ葉の病徴を画像から認識する軽量モデルを開発し、公開・自作データセットおよびエッジデバイスで性能検証しているため、植物病害状態のフェノタイピング手法が中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.