Aphid-ResNetSwin: An Image Recognition Method with Improved Attention Mechanism for Graded Identification of Myzus persicae

TL;DR

A new AI model called Aphid-ResNetSwin improves the accuracy of identifying aphid infestation levels in tobacco crops using advanced image recognition techniques.

Contribution

The novel Aphid-ResNetSwin model integrates a dual-branch hybrid architecture with a Global Convolutional Spatial Attention module for improved aphid infestation grading.

Findings

Aphid-ResNetSwin achieved 89.11% accuracy in graded recognition of Myzus persicae infestation levels.

The model outperformed baseline models like MobileNetV3 and InceptionResNetV2 in recognition accuracy.

It showed higher classification accuracy than manual identification for all infestation severity levels except healthy leaves.

Abstract

Infestation of crops by the Myzus persicae results in yield loss and quality deterioration of agricultural products. Accurately identifying M. persicae helps to develop prevention and control strategies in advance, thereby reducing related yield and quality losses. Traditional image classification methods exhibit significant limitations in terms of accuracy and robustness under complex field conditions. To address these challenges, this study proposes a novel image recognition model, Aphid-ResNetSwin, for the graded identification of tobacco aphids. This network employs a novel dual-branch hybrid neural network architecture based on Inception-ResNet-V2 and Swin Transformer, in which the Global Convolutional Spatial Attention (GCSA) module is integrated into each branch to enhance feature attention extraction. Such a design effectively improves the capability of local feature learning and recognition accuracy, thereby significantly boosting the overall recognition performance.

Myzus persicae is the most devastating piercing-sucking pest threatening tobacco production. Precise quantification of infestation severity is a prerequisite for precision pest management, making the integration of visual image analysis highly essential for efficient management. Current computer vision models in modern agriculture are primarily designed for classifying various pest species, and there is a lack of image-driven analytical tools for assessing the severity of damage inflicted by a single target pest. To supplement existing analytical tools and enable the graded recognition of tobacco aphid (M. persicae) infestation levels, we propose the Aphid-ResNetSwin model. This model is constructed by integrating a Global Channel-Spatial Attention module (a specialized attention mechanism) into the well-established InceptionResNetV2 architecture. Performance evaluation results demonstrated that the Aphid-ResNetSwin model achieved a graded recognition accuracy of 89.11%. Compared with widely adopted mainstream baseline models in computer vision, such as MobileNetV3, SwinTransformer, and InceptionResNetV2, our proposed model exhibited superior performance in recognition accuracy. Furthermore, the classification accuracy of our model for M. persicae infestation across all severity levels was significantly higher than that of manual identification, with the exception of healthy leaves. Collectively, our findings indicate that the Aphid-ResNetSwin model provides a robust tool for the graded recognition of M. persicae infestation, offering valuable model-based support for the precision control of aphids in tobacco fields.

Figures