Fuzzy-Cycle: Visible to Infrared Ship Image Translation Based on CycleGAN

Detecting targets in infrared ship images is crucial for various applications, such as maritime rescue and ship inspection. Training contemporary mainstream detection and segmentation models, however, requires large-scale, high-quality infrared ship image datasets. In open-world maritime confrontation scenarios, the complex and variable environment, scarcity of target ships, and limited acquisition conditions make it challenging to construct high-quality infrared ship image datasets. To address this issue, this paper proposes an unpaired visible-infrared ship image translation model based on CycleGAN, named Fuzzy-Cycle. The model introduces a fuzzy cycle loss to encourage the model to learn feature mapping rules between different domains while minimizing attention to unnecessary features. Additionally, the method employs U-Net as the generator network to retain shallow image features during upsampling and selects PatchGAN as the discriminator network, embedding Haar wavelet transform to enhance the discriminator's ability to analyze image frequency domain features. Experiments on the VAIS dataset demonstrate that the proposed algorithm outperforms mainstream unpaired image transfer models, showing excellent performance in the task of visible-infrared ship image translation. Furthermore, zero-shot image translation experiments on the HRSC2016 dataset prove the generalizability and reliability of the proposed method.