MAF-MixNet: Few-Shot Tea Disease Detection Based on Mixed Attention and Multi-Path Feature Fusion
Open Access
Abstract
Tea (Camellia sinensis L.) disease detection in complex field conditions faces significant challenges due to the scarcity of labeled data. While current mainstream visual deep learning algorithms depend on large-scale curated datasets. To address this, we propose a novel few-shot end-to-end detection network called MAF-MixNet that achieves robust detection with minimal annotation data. The network effectively overcomes the bottleneck of insufficient feature extraction under limited samples of existing methods, through the design of a mixed attention branch (MA-Branch) and a multi-path feature fusion module (MAFM). The former extracts contextual features, while the latter combines and enhances the local and global features. The entire model uses a two-stage paradigm to pretrain on public datasets and fine-tune on balanced subset datasets, including novel tea disease classes, anthracnose, and brown blight. Comparative experiments with six models on four evaluation metrics verified the advancement of our model. At 5-shot, MAF-MixNet achieves scores of 62.0%, 60.1%, and 65.9% in precision, nAP50, and F1 score, respectively, significantly outperforming other models. Similar superiority is achieved in the 10-shot scenario, where nAP50 is 73.8%. Our model maintains a certain computational efficiency and achieves the second fastest inference speed at 11.63 FPS, making it viable for real-world deployment. The results confirm MAF-MixNet’s potential to enable cost-effective, intelligent disease monitoring in precision agriculture.Keywords
This publication has 57 references indexed in Scilit:
- A transformer-based few-shot learning pipeline for barley disease detection from field-collected imageryComputers and Electronics in Agriculture, 2024
- Cross-Domain Few-Shot Object Detection via Enhanced Open-Set Object Detector2024
- GS-DeepLabV3+: A mountain tea disease segmentation network based on improved shuffle attention and gated multidimensional feature extractionCrop Protection, 2024
- Plant disease recognition in a low data scenario using few-shot learningComputers and Electronics in Agriculture, 2024
- Attention mechanisms in computer vision: A surveyComputational Visual Media, 2022
- Estimation of tea leaf blight severity in natural scene imagesPrecision Agriculture, 2021
- End-to-End Object Detection with Transformers2020
- Tea: A Worthwhile, Popular Beverage Crop Since Time Immemorial2019
- A low shot learning method for tea leaf’s disease identificationComputers and Electronics in Agriculture, 2019
- SSD: Single Shot MultiBox Detector2016