CNN-LRP: Understanding Convolutional Neural Networks Performance for Target Recognition in SAR Images

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Synthetic Aperture Radar (SAR) image target recognition is complex, often requiring extensive pre-processing.
• Deep learning, especially Convolutional Neural Networks (CNNs), offers potential but suffers from a "black box" problem, hindering error analysis.
• Existing visualization methods like Layer-wise Relevance Propagation (LRP) are not optimally suited for CNNs in SAR applications.

Purpose of the Study:

• To develop a novel LRP algorithm specifically for understanding CNN performance in SAR image target recognition.
• To provide a clear visual explanation of CNNs' recognition mechanisms by highlighting input contributions.
• To overcome the limitations of existing LRP methods when applied to CNNs.

Main Methods:

• A novel LRP algorithm is proposed, tailored for CNNs used in SAR image analysis.
• The method derives a concise form of the correlation between CNN layer outputs and subsequent layer weights.
• The algorithm visualizes positive and negative contributions within input SAR images to CNN classification.

Main Results:

• The proposed LRP method effectively visualizes the inner workings of CNNs for SAR target recognition.
• It provides clear insights into which parts of the SAR image influence the classification outcome.
• Experimental results confirm the proposed method's superiority over common LRP techniques.

Conclusions:

• The novel LRP algorithm enhances the interpretability of CNNs in SAR image target recognition.
• This approach facilitates better error analysis and understanding of CNN decision-making.
• The method offers a valuable tool for researchers in SAR image processing and computer vision.