Abstract:

To address the challenge of lacking effective metrological evaluation methods for the output values of deep learning object detection algorithms due to their lack of interpretability, this study constructs a technical framework for algorithm traceability from a metrological perspective. Using ship detection in Synthetic Aperture Radar (SAR) images as a typical application scenario, the definition and uncertainty sources of algorithm traceability are clarified, and a traceability technical path based on standard datasets together with a continuous comparison chain is proposed. To meet the metrological requirements of algorithm traceability, a quality evaluation indicator system and standardized testing methods for standard datasets are established. This study has achieved quantitative evaluation and traceability of the performance metrics generated by SAR image-based ship detection algorithms. It provides a benchmark for the reliable evaluation of artificial intelligence algorithms and is of great significance for advancing the development of a standardized evaluation system.

Key words: algorithm traceability, standard datasets, deep learning, quality assessment