Ship Target Detection in Optical Remote Sensing Images Based on Multiscale Feature Enhancement.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Ship target detection in optical remote sensing images (ORSIs) is challenging due to the multiscale nature of targets.
• Existing deep learning methods struggle with low accuracy in detecting ships of various sizes.

Purpose of the Study:

• To improve the accuracy of multiscale ship target detection in ORSIs.
• To propose a novel ship detection algorithm based on enhanced multiscale feature extraction using YOLOv4.

Main Methods:

• Introduced Mixed Inverted Residual Blocks (MIRes) by integrating improved mixed convolution into inverted residual blocks.
• Replaced standard residual blocks in the backbone network's CSP module with MIRes to enhance multiscale feature extraction.
• Developed Small Scale Feature Enhancement Modules (SFEM) and Middle Scale Feature Enhancement Modules (MFEM) to boost feature information in low- and middle-level feature maps.

Main Results:

• The proposed algorithm achieved 79.55% and 90.70% accuracy on the LEVIR-ship and NWPU VHR-10 datasets, respectively.
• Demonstrated accuracy improvements of 3.25% and 3.56% compared to the standard YOLOv4 algorithm.
• Effectively enhanced feature representation for multiscale ship targets.

Conclusions:

• The proposed multiscale feature enhancement strategy significantly improves ship detection accuracy in ORSIs.
• The integration of MIRes and SFEM/MFEM modules offers a robust solution for challenges posed by multiscale targets.
• The enhanced YOLOv4-based algorithm shows superior performance for ship detection in complex remote sensing scenarios.