Image compression is a topic of significant interest as it reduces file sizes in stored data. In this paper, we propose a model that achieves multiple levels of compression, thereby minimizing the storage space required for images, which typically consume substantial amounts of data due to their size and resolution. We combine an image downscaling and upscaling model with an image compression model. By leveraging convolutional techniques to identify image features, we can effectively reduce the size of the image through downscaling and subsequently upscaling it. Additionally, we employ entropy image compression and arithmetic encoding to compress and reconstruct the image while preserving its lossless data. Through experimentation with the Kodak dataset, we observed that our proposed model achieved a compression rate of 96.92%, significantly reducing the data needed for file storage. Moreover, our reconstructed images attained a standardized measure with a signal-to-noise ratio of 33.10 dB and a structural similarity of 0.9219. Notably, the perceptual quality of the images, including intricate details, remained intact to the human eye.
Paper
Full text
Compressed Image Super Resolution using Convolutional Neural Network
Semantic Scholar · Computer Science · 2023
Abstract
Image compression is a topic of significant interest as it reduces file sizes in stored data. In this paper, we propose a model that achieves multiple levels of compression, thereby minimizing the storage space required for images, which typically consume substantial amounts of data due to their size and resolution. We combine an image downscaling and upscaling model with an image compression model. By leveraging convolutional techniques to identify image features, we can effectively reduce the size of the image through downscaling and subsequently upscaling it. Additionally, we employ entropy image compression and arithmetic encoding to compress and reconstruct the image while preserving its lossless data. Through experimentation with the Kodak dataset, we observed that our proposed model achieved a compression rate of 96.92%, significantly reducing the data needed for file storage. Moreover, our reconstructed images attained a standardized measure with a signal-to-noise ratio of 33.10 dB and a structural similarity of 0.9219. Notably, the perceptual quality of the images, including intricate details, remained intact to the human eye.