Large kernel convolution YOLO for ship detection in surveillance video.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current ship detection models suffer from numerous hyperparameters, leading to suboptimal recognition accuracy and imprecise boundary regression.
• Existing methods often face challenges in fusing regression and classification tasks within coupled detection heads.

Purpose of the Study:

• To design an efficient and accurate one-stage ship detection model, Lk-YOLO (Large Kernel Convolutional YOLO), utilizing an anchor-free approach.
• To enhance feature extraction capabilities and improve the precision of bounding box regression.

Main Methods:

• Incorporated large kernel convolutions into the backbone network's residual modules for superior feature extraction.
• Decoupled the detection head into two branches to optimize classification and regression tasks independently.
• Implemented an anchor-free algorithm and an improved sample matching strategy to address hyperparameter complexity and class imbalance.

Main Results:

• The Lk-YOLO model achieved a mean Average Precision (mAP@50) of 97.7% and mAP@.5:.95 of 78.4%.
• Demonstrated superior accuracy and robustness compared to existing ship detection models.
• Eliminated the need for anchor hyperparameter design, reducing computational complexity.

Conclusions:

• The proposed Lk-YOLO model offers a simplified yet highly effective solution for ship detection.
• The anchor-free design and enhanced feature extraction contribute to improved detection efficiency and robustness.
• This approach provides a strong foundation for future advancements in maritime surveillance and object detection.