ASP-Det: Toward Appearance-Similar Light-Trap Agricultural Pest Detection and Recognition

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Intelligent agriculture faces challenges in automatic pest detection due to visual similarities between pests in 2D images.
• The appearance-similarity problem in pest detection is characterized by texture and scale similarities, hindering accurate recognition.

Purpose of the Study:

• To introduce a new task, the Appearance Similarity Pest Detection (ASPD) task, specifically for agricultural pest detection.
• To propose novel quantitative metrics (Multi-Texton Histogram and Object Relative Size) for texture and scale similarity.
• To develop an effective detection method (ASP-Det) and a task-specific dataset (PestNet-AS) for the ASPD task.

Main Methods:

• Developed two novel metrics: Multi-Texton Histogram (MTH) for texture similarity and Object Relative Size (ORS) for scale similarity.
• Introduced the ASP-Det method, incorporating a Pairwise Self-Attention (PSA) mechanism and Non-Local Modules for texture similarity.
• Integrated a Skip-Calibrated Convolution (SCC) module to address scale variations and re-calibrate feature maps, forming a one-stage anchor-free detection framework.

Main Results:

• Created and annotated a new dataset, PestNet-AS, tailored for the ASPD task.
• The proposed ASP-Det method demonstrated strong performance as a baseline for the ASPD task on the PestNet-AS dataset.
• The combination of PSA-Non Local and SCC modules effectively addressed texture and scale similarity challenges in pest detection.

Conclusions:

• The ASPD task and the proposed ASP-Det method provide a significant advancement in agricultural pest detection.
• The novel metrics and modules offer a quantitative approach to understanding and solving appearance similarity issues in pest recognition.
• ASP-Det serves as a robust baseline, paving the way for more accurate and efficient intelligent agriculture systems.