FsDAOD: Few-shot domain adaptation object detection for heterogeneous SAR image

Abstract

Heterogeneous Synthetic Aperture Radar (SAR) image object detection task with inconsistent joint probability distributions is occurring more and more frequently in practical applications. In which the small sample of data scarcity is becoming an urgent problem for researchers. Therefore, this paper proposes a novel few-shot domain adaptation object detection (FsDAOD) method based on Faster Region Convolutional Neural Network baseline to cope with the above problem. Firstly, employing the foundational structure of the existing baseline method, a novel mutual information loss function is introduced that prompts the neural network to extract domain-specific knowledge. This strategic approach encourages distinctive levels of confidence in individual predictions while fostering overall diversity. Given that performance can be easily over-fitted with a restricted number of observed objects if feature alignment strictly adheres to conventional methods, the set of source instances are initially categorized into two groups: target domain-easy set and target domain-hard set. Subsequently, asynchronous alignment is performed between the target-hard domain set of the source instances and the extended dataset of the target instances to achieve effective supervised learning. It is then asserted that confidence-based sample separation methods can improve detection efficiency by adjusting the model to prioritize the identification of more easily detected objects, but this may lead to incorrect decisions for more challenging instances. Extensive experiments on FsDAOD on heterogeneous satellite-borne SAR image datasets have been conducted, and the experimental results have demonstrated that the detection rate of the proposed method exceeds the existing state-of-the-art methods by 5%.