A Novel Deep Learning-based Image Compression Approach for Achieving High Fidelity and Efficient Storage

Efficient image compression is essential for storage and transmission in several applications, including medical imaging, remote sensing, and multimedia. Traditional compression approaches struggle to preserve crucial details while maintaining high compression ratios, especially when specific portions of a picture contain more significant information than others. Recent advancements in neural image compression (NIC) and deep learning-based segmentation have enabled more adaptive compression strategies by distinguishing between regions of interest (ROI) and non-ROI areas. Despite these advancements, existing methods lack bitrate allocation flexibility, compromising image integrity. To address these challenges, this work proposes a hybrid neural image compression (NIC) approach that integrates multi-scale ROI segmentation, adaptive bitrate allocation, and adversarial robustness testing. For Multiscale ROI Segmentation: For ROI and NonROI Regions, DeepLabV3+ and Segformer application-based segmentation models are used. Lossless compression is performed on ROI regions, and lossy compression (NIC models) is used on non-ROI regions. The method is evaluated against multiple datasets: CIFAR-10, STL-10, medical images, satellite images, and MRI-CT scans. We benchmark our compression performance with PSNR, SSIM, LPIPS, and VIF metrics, demonstrating the effectiveness of our proposed approach in preserving important information while achieving high compression ratios.