Identification of plant leaf diseases by deep learning based on channel attention and channel pruning

• DOI: 10.3389/fpls.2022.1023515
• Publication Date: 2022-11-10T10:09:53Z
• Authors: Riyao Chen; Haixia Qi; Yu Liang; Mingchao Yang
• Keywords: Pruning; Pooling; Identification; FLOPS

Plant diseases cause significant economic losses and food security in agriculture each year, with the critical path to reducing being accurate identification timely diagnosis of plant diseases. Currently, deep neural networks have been extensively applied disease identification, but such approaches still suffer from low accuracy numerous parameters. Hence, this paper proposes a model combining channel attention pruning called CACPNET, suitable for common species. The mechanism adopts local cross-channel strategy without dimensionality reduction, which is inserted into ResNet-18-based that combines global average pooling max effectively improve features' extracting ability leaf Based on model's optimum feature extraction condition, unimportant channels are removed reduce parameters complexity via L1-norm weight compression ratio. CACPNET public dataset PlantVillage reaches 99.7% achieves 97.7% peanut dataset. Compared base ResNet-18 model, floating point operations (FLOPs) decreased by 30.35%, 57.97%, size 57.85%, GPU RAM requirements 8.3%. Additionally, outperforms current models considering inference time throughput, reaching 22.8 ms/frame 75.5 frames/s, respectively. results outline appealing deployment edge devices efficiency precision detection.