Patch-Attention GAN: Image Translation using BiFormer and Attention Framework

Image-to-image translation has long been recognized as a crucial undertaking within the field of computer vision, owing to its broad applications in domain adaption. Existing methods differ in Generative Adversarial Network framework and model structure. Attention-Guided training framework absorbs the input image’s semantic information based on the translation task. However, it ignores the local context for content extraction and lacks mask comparison in the loss function. The attention mechanism is widely regarded as an effective approach for addressing contextual challenges in machine learning models. By augmenting the model’s capacity to capture long-range dependencies and concentrate on the most salient aspects of the input sequence, the attention mechanism significantly enhances its contextual understanding. Bi-level routing attention stands out among other mechanisms for its computation speed and memory cost. In this paper, we propose an image translation algorithm called Patch-Attention Generative Adversarial Network that combines foreground-background separation processing, bi-level routing attention, and content loss. Our proposed method has been extensively evaluated on diverse datasets, yielding compelling experimental results. These findings demonstrate the substantial performance boost achieved by our approach in the domain of image translation. The generated image is more clarified and understandable compared with other existing approaches.