In a data-driven world, every type of data requires a physical drive in order to be stored. The number of photos, videos, and other comparable data kinds that are shared online has increased dramatically. Efficient storage and transmission of data has made image compression an important topic of research. This paper proposes two recent deep learning-based image compression frameworks to get encouraging advancements and better rate-distortion efficiency when compared to classical or traditional image compression schemes. Autoencoders and vision transformers eliminate the limitations of fixed mathematical models because they deal with more advanced representations of image content, capturing complex patterns and global properties of an image. This paper work has empirically investigated the impact of two deep learning-based frameworks based on image compression and also, we have compared them by evaluating their mean square error, peak signal-to-noise ratio, and compression ratio using CIFAR-10 datasets to find out which deep learning approach gives good promising results maintaining image quality to expand the scope of image compression. Experimental results reveal that autoencoders, with their convolutional structure, provide a balanced approach in reducing image dimensions and also preserves a reasonable image quality. Further, they yield reasonably good peak signal-to-noise ratio (PSNR), low mean square errors (MSE), and high ratio of compression which make them relatively efficient for compression. In addition, they require low computational cost and complexity as compared to other methods. On the other hand, vision transformers (ViT) yield very high peak signal-to-noise ratio (PSNR) values but due to its aggressive compression it degrades the image quality. Vision transformers are comparatively expensive computationally and also its complex training architecture may not make it resource friendly.