Pest recognition based on multi-image feature localization and adaptive filtering fusion.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate pest recognition is vital for effective pest control but faces challenges due to appearance variations, data quality, and complex environments.
• Existing single-image recognition models struggle with accuracy, with the highest reaching only 75% on the IP102 dataset.

Purpose of the Study:

• To propose and evaluate a novel multi-image fusion recognition method to improve pest identification accuracy in practical agricultural settings.
• To leverage farmers' easy access to multiple images of the same pest for enhanced recognition.

Main Methods:

• Utilizes convolutional neural networks (CNNs) to extract feature maps from multiple images of the same pest.
• Employs an effective feature localization module (EFLM) to identify and localize pest features.
• Integrates an adaptive filtering fusion module (AFFM) with an attention mechanism for feature selection and fusion, followed by a soft voting (SV) classifier for final categorization.

Main Results:

• The proposed method achieved high recognition accuracies: 73.9% on IP102, 99.8% on D0, and 99.7% on ETP using single images.
• Multi-image fusion significantly boosted accuracy to state-of-the-art levels: 96.1% (5 images) on IP102, 100% (2 images) on D0, and 100% (2 images) on ETP.
• The developed web application demonstrates practical application for farmers, aiding reliable pest identification.

Conclusions:

• The multi-image fusion approach effectively overcomes limitations of single-image pest recognition.
• The method's high accuracy and practical applicability support the advancement of smart agriculture and farmer assistance.
• The integration of EFLM, AFFM, and SV modules contributes to the superior performance of the fusion model.