Cascaded Aggregation Convolution Network for Salient Grain Pests Detection.

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Pest infestations in stored grains cause significant economic losses.
• Accurate detection of small grain pests is difficult due to their size, variability, and background clutter.
• Salient object detection enhances pest identification by focusing on visually prominent features.

Purpose of the Study:

• To develop a novel deep learning model for accurate salient pest detection in stored grains.
• To improve pest detection accuracy by mimicking human visual attention mechanisms.
• To enable effective pest control and infestation severity assessment in grain storage.

Main Methods:

• Proposed a Cascaded Aggregation Convolution Network (CACNet) using VGG16 as the backbone.
• Implemented reverse cascade feature aggregation, feature enhancement, and feature aggregation operations.
• Curated the GrainPest dataset and utilized the MSRA-B dataset for validation.

Main Results:

• CACNet achieved a structure S-measure of 91.9% and 90.9% on the datasets.
• The model obtained a weighted F-measure of 76.4% and 91.0%.
• Demonstrated superior performance compared to traditional and deep learning-based saliency detection methods.

Conclusions:

• The developed CACNet effectively detects small-scale grain pests with high accuracy.
• This technology shows significant potential for real-time pest monitoring and infestation assessment in grain storage.
• The approach is promising for pest prevention and control in agriculture and forestry.