We propose an approach to enhance the local implicit neural network representation for decoding high-quality images. The JPEG algorithm quantizes coefficients in the discrete cosine frequency domain into a small set to achieve high compression ratios....
We propose an approach to enhance the local implicit neural network representation for decoding high-quality images. The JPEG algorithm quantizes coefficients in the discrete cosine frequency domain into a small set to achieve high compression ratios. Therefore, quality degradation is inevitable in traditional JPEG decoding methods. To improve the quality of decoded JPEG images, we introduce a continuous cosine coefficient extractor into the network. Through learning as a function of interval coordinates in JPEG, the proposed network can restore overall quality coefficients. This approach takes distorted cosine coefficients as input, restores the quantized coefficients, and applies them to an implicit neural network to decode high-quality images. As a result, the proposed method achieves state-of-the-art performance in terms of compressed image restoration for various quality coefficients with a single model.