During the past decades, applying deep learning technologies on fashion industry are increasingly the mainstream. Due to the different gesture, illumination or self-occasion, it is hard to directly utilize the clothes images in real-world applications. In this paper, to handle this problem, we present a novel multi-stage, category-supervised attention-based conditional generative adversarial network by generating clear and detailed tiled clothing images from certain model images. This newly proposed method consists of two stages: in the first stage, we generate the coarse image which contains general appearance information (such as color and shape) and category of the garment, where a spatial transformation module is utilized to handle the shape changes during image synthesis and an additional classifier is employed to guide coarse image generated in a category-supervised manner; in the second stage, we propose a dual path attention-based model to generate the fine-tuned image, which combines the appearance information of the coarse result with the high-frequency information of the model image. In detail, we introduce the channel attention mechanism to assign weights to the information of different channels instead of connecting directly. Then, a self-attention module is employed to model long-range correlation making the generated image close to the target. In additional to the framework, we also create a person-to-clothing data set containing 10 categories of clothing, which includes more than 34 thousand pairs of images with category attribute. Extensive simulations are conducted, and experimental result on the data set demonstrates the feasibility and superiority of the proposed networks.

在过去的几十年中，将深度学习技术应用于时尚行业越来越成为主流。由于手势，照明或自身场合的不同，在实际应用中很难直接利用衣服图像。在本文中，为了解决这个问题，我们通过从某些模型图像生成清晰且详细的平铺衣服图像，提出了一种新颖的多阶段，类别监督的基于注意力的条件生成对抗网络。这种新提出的方法包括两个阶段：在第一阶段，我们生成包含常规外观信息（例如颜色和形状）和服装类别的粗略图像，利用空间变换模块来处理图像合成过程中的形状变化，并使用附加的分类器来指导以类别监督的方式生成的粗糙图像；在第二阶段，我们提出了一种基于双路径注意力的模型来生成微调图像，该模型将粗糙结果的外观信息与模型图像的高频信息相结合。详细地，我们介绍了通道注意机制，以将权重分配给不同通道的信息，而不是直接连接。然后，采用自我关注模块对远距离相关性进行建模，使生成的图像接近目标。除了框架之外，我们还创建了一个包含10种衣服的“人到衣服”数据集，其中包括超过34,000对具有类别属性的图像。进行了广泛的仿真，数据集上的实验结果证明了所提出网络的可行性和优越性。