Small Sample Set Inshore Ship Detection From VHR Optical Remote Sensing Images Based on Structured Sparse Representation

• Main Authors: Yin Zhuang, Lianlin Li, He Chen
• Format: Article
• Language: English
• Published: IEEE 2020-01-01
• Series: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
• Subjects: Inshore ship detection; optical remote sensing; sparse representation (SR); small sample set; very high resolution (VHR)
• Online Access: https://ieeexplore.ieee.org/document/9076854/

Inshore ship detection from very high resolution (VHR) optical remote sensing images has been playing a critical role in various civil and military applications. However, it brings up an important challenge, which is difficult to complete effective and robust feature extraction when valid inshore ship training sample acquired is limited, and the severe imbalance problem exists of positive and negative samples. In order to tackle the abovementioned difficulties, the structured sparse representation model (SSRM) is proposed to achieve inshore ship detection in more effectively and robustly way by circumstances of the small sample set. Here, SSRM has two steps that include inshore ship region proposal (RP) and orientation prediction (OP). Related to the RP process, the error matrix embedded in SSRM not only prevents to build the high-dimension background subdictionary and imbalance problem of positive and negative samples, but also achieves an effective intraclass robustness description of inshore ships and background. For the OP stage, the low-rank constraint of common sharing atoms in SSRM can make inshore ship direction be extracted by their sparse coding. In addition, based on RP and OP guidance, the proposed comprehensive structure voting can achieve an accurate contour detection of inshore ships. Finally, several experimental results employ that Google Earth service, HRSC 2016, and DOTA datasets proved the effectiveness of the proposed method. The results show that proposed inshore ship detection method can provide approximately 83.7% Recall and 72.3% Precision by using only over 100 positive training samples, which outperforms the state of the art methods.