Semantic learning for image compression (SLIC)

Abstract

Image compression is an ever-evolving problem with different approaches coming into prominence at different times. Data analysts are still contemplating about the best approach with compression ratio, visual quality and complexity of the architecture as the main criteria of evaluation. Most recent one is the advancements in machine learning and its applications on image classification. Researchers are currently exploring the possibility of using this advanced computing power to enhance the quality of images with the availability of large datasets. Here, we present a deep learning Convolutional neural network (CNN) implemented and trained for image recognition and image processing tasks, to understand semantics of an image. We achieve higher visual quality in lossy compression by increasing the complexity of an encoder by a slight margin to generate semantic maps of image. The semantic maps are tailored to make the encoder content-aware for a given image. In the proposed work, we present a novel architecture developed specifically for image compression, which generates a semantic map for salient regions so that they can be encoded at higher quality as compared to background regions. Experiments are conducted on the Kodak PhotoCD dataset and the results are compared with other state-of-the-art models. Results of the conducted experiment reveal an increase in compression ratio by more than 17% compared to the referred state-of-the-art model while maintaining the visual quality.