Float-DEIM: An enhanced transformer model for small floating waste detection

Area of Science:

• Environmental Science
• Computer Vision
• Remote Sensing

Background:

• Global water pollution necessitates accurate floating waste identification for effective water environment management.
• Detecting small targets in complex water surfaces is challenging due to limited pixel data and background interference.

Purpose of the Study:

• To develop a high-precision floating waste detection model for complex water environments.
• To address the challenges of small target detection in water pollution monitoring.

Main Methods:

• Proposed the Partial Efficient Multi-Scale Attention (PEMA) mechanism for balanced multi-scale spatial modeling and feature preservation.
• Introduced the Progressive Partial Convolution Downsample (PPCD) module to mitigate information loss during downsampling.
• Developed a Dual-stage Focusing Pyramid Network (DFPN) for enhanced feature discrimination and localization accuracy.

Main Results:

• The Float-DEIM model achieved a 2.7% improvement in small target detection accuracy (AP_s^test) on the IWHR_AI_Label_Floater_V1 dataset compared to the baseline.
• Demonstrated Float-DEIM's adaptability and effectiveness across different environmental scenes through generalization experiments.

Conclusions:

• Float-DEIM offers a robust technical solution for automatic floating waste detection in challenging aquatic environments.
• The proposed PEMA, PPCD, and DFPN modules effectively enhance small target detection capabilities.