SFL-MobileNetV3: A lightweight network for weed recognition in tropical cassava fields in China

Abstract

Weed interference significantly impacts on cassava yields in tropical farmlands in Hainan, China. Incomplete datasets and technological limitations in lightweight recognition hinder the real-time and accurate identification of weeds in cassava fields. In this study, TropCropW38, a tropical weed dataset designed for cassava yield analysis, was developed and comprises 7,616 images representing 38 distinct weed species. Based on TropCropW38, an enhanced lightweight model, SFL-MobileNetV3, was proposed. SFL-MobileNetV3 builds upon MobileNetV3-Large and incorporates a Simple Attention Module (SimAM) within the Bneck module to reduce parameter count and mitigate background interference. A Feature Pyramid Network (FPN) is integrated into the backbone, combined with a feature fusion strategy to produce enriched feature representations that enhance the model’s ability to extract weed images of varying sizes and perspectives in complex environments. To address class imbalance and complex background interference in the dataset, a composite loss function combining cross-entropy loss, focal loss, and label smoothing was proposed. Comparative experiments were conducted with ten classical and lightweight models to evaluate the SFL-MobileNetV3′s performance. Experimental results indicate that SFL-MobileNetV3 achieves a recognition accuracy of 94.34 % on the TropCropW38 dataset, a 7.46 % improvement over MobileNetV3-Large. The model also demonstrates improvements of 7.86 %, 9.63 %, and 9.71 % in accuracy, recall, and F1-score, as well as a 5.20 MB reduction in model size, with a computational complexity of 0.60 GFLOPs. With its high accuracy, low computation complexity, lightweight architecture, and robust performance, the enhanced model shows promise for real-time weed recognition in tropical farmlands, especially for deployment on mobile or embedded systems

Topics & Concepts