"SpatialMix-JSCC"

"\u4e09\u79cd\u6570\u636e\u96c6\uff1aDIV2K\u3001CelebA\u3001AFHQDIV2K DatasetDIV2K (DIVerse 2K resolution high-quality images) is a benchmark dataset widely adopted in image super-resolution (SR) and restoration research, initially introduced for the NTIRE 2017 and NTIRE 2018 challenges at CVPR and later extended for PIRM 2018 at ECCV. It comprises 1,000 high-resolution 2K images (2048\u00d71080 pixels minimum on one axis) with diverse content spanning natural scenes, architecture, landscapes, and objects. The dataset is systematically partitioned into 800 training images, 100 validation images, and 100 testing images. For SR tasks, DIV2K provides paired low-resolution (LR) and high-resolution (HR) images generated using various degradation models (bicubic downsampling, unknown complex downgrading) and scale factors (\u00d72, \u00d73, \u00d74). Its standardized structure and realistic degradations have established it as the de facto standard for evaluating SR algorithms, driving advancements in single-image super-resolution, image compression, and restoration techniques.CelebA DatasetThe Large-scale CelebFaces Attributes (CelebA) dataset, developed by the Multimedia Laboratory (MMLAB) at the Chinese University of Hong Kong, is a foundational resource for face-related computer vision research. It contains 202,599 celebrity face images across 10,177 unique identities, making it one of the largest publicly available face attribute datasets. Each image is annotated with rich metadata, including 5 facial landmark coordinates (eyes, nose, mouth) and 40 binary attributes (e.g., \"Smiling,\" \"Wearing Glasses,\" \"Bald,\" \"Male\"). The dataset exhibits significant diversity in pose variations, background clutter, ages, ethnicities, and accessories, enabling the training of robust, generalizable models. CelebA is extensively used for tasks such as face attribute recognition, facial landmark localization, face detection, and has become a standard benchmark for evaluating generative models in face synthesis and manipulation. Its comprehensive annotations also support research in fair face recognition and bias mitigation.AFHQ DatasetAnimal Faces-HQ (AFHQ) is a high-quality dataset designed specifically for animal face analysis and cross-domain image translation research, introduced alongside StarGAN v2 at CVPR 2020. It consists of 15,000 512\u00d7512 pixel images divided equally into three distinct domains: cats, dogs, and wildlife (including lions, tigers, pandas, and other exotic species). Each domain features at least eight different breeds or subspecies, ensuring substantial intra-domain diversity. The dataset follows a standard split of 4,500 training images and 500 test images per domain, with precise image alignment focusing on animal facial features. AFHQ addresses the scarcity of high-quality animal face datasets with controlled splits, making it ideal for training and evaluating generative adversarial networks (GANs) for unpaired image-to-image translation, domain adaptation, and few-shot learning tasks. Its clean, high-resolution images have also been used for animal face detection, classification, and attribute recognition research."