CLIP Text-Guided Panoramic Image Style Transfer Based on DLKA Attention
Yue Wang, Boning Zhang, Wujian Ye, Yijun Liu, Dongjun Tan · 2024
The text-guided image style transfer based on CLIP can help users freely generate images with different styles without relying on reference images. On one hand, this method eliminates the need for large style image datasets to train the model; on the other hand, it allows for more precise specification of the style features to be transferred. Considering the advantage of panoramic images having a wide field of view, this paper integrates DLKA attention mechanism with CLIP text-guided style transfer to design a text-guided panoramic image style transfer framework. This enhances the style transfer network's ability to recognize and extract features from the spatially distorted parts of panoramic images. Experimental results show that the proposed method achieves significant improvements in SSIM and PSNR indicators. Compared to the original method and other attention mechanisms, this method improves SSIM by about 0.1-0.3 and PSNR by 1-4dB.