Fine-Grained Recognition Framework For Ship Targets In SAR Images Based On Component Architecture

Applying deep learning network models in SAR image processing has significantly improved speed of target recognition performance. Nevertheless, due to the SAR imaging principle distinct from optical images, these models struggle to distinguish strong scattering characteristics from clutter, resulting in a high false alarm rate in applications of target recognition. To ensure correctness practically, the expert analysis for discern various structures is strongly necessary for fine-grained recognition after possessing through integrating intelligent models, with extra time and manpower cost. This study introduces a fine-grained recognition framework for ship targets in SAR images based on component architecture. A ship targets dataset from Gaofen-3 SAR images, with three classified ship types is designed to train the model, including meticulous analysis and labeling for target categorization, axis orientation, and key component localization. Features of different dimensions extracted via pre-trained ResNet-50 to construct a weight dossier, and a multi-task joint detection mode are embedded into the framework to integrate the experts’ ship recognition method built on component architecture into the intelligent network model. This yields an accuracy rate surpassing 89% for ship identification, eclipsing conventional intelligence models by a significant margin. The framework successfully bolsters the ability of target awareness affiliated with SAR ocean monitoring.