A Feature-Map-Based Method for Explaining the Performance Degradation of Ship Detection Networks

Abstract

The unknowability of the inner workings limits the magnitude of performance improvement of ship target detection networks in synthetic aperture radar (SAR) images under Gaussian noise. However, none of the existing interpretation methods explain the phenomenon of network changes under noise. The feature map can visually reflect the changes in image delivery in the network, and some metrics can quantitatively characterize the degree of network performance degradation in a noise environment. So, in this article, we propose a comprehensive analysis method that integrates texture and brightness features of the internal feature map of the network to clarify the change process of target features under Gaussian noise. First, we analyzed the degradation of three target detection networks under different levels of Gaussian noise; then, the feature maps of four convolution layers were sampled and visualized for qualitative analysis; finally, the texture and brightness features were extracted for quantitative characterization of the feature amount changes. We experimentally validated the method on publicly available SSDD radar datasets. The networks were extremely sensitive to Gaussian noise, and the mean Average Precision decreased by up to 96.3%. The angular second moment and entropy texture feature values of the feature map could drop and rise 59.10% and 97.81%, respectively, while the brightness value could increase up to 100.92%. This indicates that noise changes the structure of feature maps and reduces the amount of effective information.

Keywords

• Interpretability
• ship detection
• synthetic aperture radar
• visualization