Learning feature fusion for target detection based on polarimetric imaging.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Polarimetric imaging captures rich information about material properties and surface structures.
• Traditional target detection methods struggle with complex scenes and subtle features.
• Feature fusion offers a promising avenue for enhancing target detection performance.

Purpose of the Study:

• To develop and evaluate a novel polarimetric imaging processing method for target detection.
• To leverage feature fusion techniques with convolutional neural networks (CNNs) for improved target recognition.
• To create a comprehensive dataset for training and validating the proposed method.

Main Methods:

• A feature fusion strategy was designed and implemented using CNNs, integrating four polarization orientation images into a single feature map.
• A new dataset was generated, including original images, polarization images, ground truth masks, and bounding boxes.
• The proposed method was compared against conventional deep learning approaches for target detection.

Main Results:

• The proposed polarimetric imaging method achieved a mean average precision (mAP) of 0.80.
• The method demonstrated a miss rate (MR) of 0.09, indicating superior performance.
• Experimental results showed significant improvements over conventional deep learning methods.

Conclusions:

• Feature fusion in polarimetric imaging, powered by CNNs, substantially enhances target detection capabilities.
• The developed method offers a robust and effective solution for target detection tasks.
• The findings suggest a promising direction for future research in advanced imaging and detection systems.