ERDNet: Efficient Ship Object Detection in Haze Environment.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Ship object detection is challenging in hazy environments due to reduced visibility.
• Current convolutional neural networks (CNNs) struggle to achieve satisfactory detection results in such conditions.

Purpose of the Study:

• To propose ERDNet, an end-to-end dual-branch network designed to enhance ship detection accuracy in hazy weather.
• To address the limitations of existing methods in handling low-quality images and multi-scale targets.

Main Methods:

• A two-branch feature extraction network with complementary attentional fusion to enhance object features.
• ERPSA-PAN, a feature pyramid fusion structure for effective context aggregation and background interference suppression.
• Spatial-Frequency Fusion Block (SFFB) module in the detection head for improved multi-scale target detection.
• A robust haze loss function to manage varying degrees of haze.

Main Results:

• The proposed ERDNet significantly improves ship detection accuracy in hazy conditions.
• ERDNet demonstrates superior performance compared to state-of-the-art (SOTA) methods.
• New datasets, Hazy-SeaShips and Hazy-Boats, were introduced to facilitate research in hazy ship detection.

Conclusions:

• ERDNet offers a robust solution for ship detection in challenging hazy environments.
• The developed network effectively enhances feature extraction, context aggregation, and multi-scale detection.
• The introduction of new datasets aids in advancing research in this domain.