NARV: An Efficient Noise-Adaptive ResNet VAE for Joint Image Compression and Denoising

Image compression and denoising are two fundamental problems in image processing and have many real-world applications. In recent years, learning-based compression methods have achieved promising rate-distortion performance. However, most learning-based compression methods are trained with clean-to-clean image pairs in a noise-unaware manner, resulting in bits misallocation when input images contain noticeable noises. To tackle this problem, we propose an efficient end-to-end method, named Noise-Adaptive ResNet VAE (NARV), which is capable of handling both clean and noisy input images. Experimental results show that NARV achieves better rate-distortion performance under various noise levels while maintaining a small model size and fast processing time for practical implementation. The code of our method is publicly accessible at https://github.com/Eventhyn/NARV