ESAK-YOLO: multi-scale attention enhanced object detection for floating waste in complex water scenes

Area of Science:

• Computer Vision
• Environmental Monitoring
• Machine Learning

Background:

• Accurate detection of floating debris in aquatic environments is challenging due to occlusion, scale variations, and background interference.
• Existing object detection models struggle with spatial feature modeling, scale adaptation, and capturing irregular object shapes.

Purpose of the Study:

• To develop an enhanced object detection model for accurately identifying floating debris in complex aquatic environments.
• To address limitations in multi-scale feature representation, scale variation perception, and irregular shape adaptability in current models.

Main Methods:

• Proposed an efficient multi-scale attention (EMA) module for improved semantic modeling via grouped feature modeling and cross-spatial fusion.
• Introduced a selective kernel attention (SKA) mechanism for dynamic receptive field adjustment and scale-variant feature capture.
• Developed an arbitrary kernel convolution (AKConv) module to enhance adaptability to irregular debris morphology.

Main Results:

• The integrated ESAK-YOLO model demonstrated significant improvements over baseline YOLOv7, increasing mean Average Precision (mAP) by 8.14%.
• ESAK-YOLO achieved a maximum improvement of 13.78% compared to five other state-of-the-art models.
• The proposed modules effectively enhance spatial feature modeling, scale adaptation, and shape capture for debris detection.

Conclusions:

• ESAK-YOLO offers a superior and effective solution for floating debris detection in challenging aquatic conditions.
• The novel EMA, SKA, and AKConv modules significantly contribute to the model's enhanced performance.
• The findings highlight the potential of ESAK-YOLO for environmental monitoring and debris management applications.