Semantic-Aware Image Compression with Masking and Generative Reconstruction for Edge Deployment

Abstract

Surveillance cameras are increasingly deployed at the network edge, continuously streaming high-resolution imagery to remote servers for processing and analysis. This becomes an issue in remote locations served by network links where the available bandwidth is severely constrained. To fit images within such limits, aggressive compression is required. Standard codecs such as JPEG, however, treat all image regions equally, degrading vehicles, pedestrians and road markings too, alongside the sky and vegetation that carry no useful information. The result is that the information from the regions most critical to downstream tasks is lost to visual artifacts. This thesis presents Masking-based Semantic-aware Image Compression (MSIC), a framework that solves this problem without requiring any changes to the existing codecs. On the sender side, a semantic segmentation model guides a patch-level masking step that zeroes out low-importance regions before JPEG encoding. Zero-valued patches encode to nearly zero DCT coefficients and therefore can be encoded with fewer bits, resulting in significant compression savings. On the receiver side, an autoencoder-based inpainting neural network fills in the masked regions from surrounding context. This design is intentionally asymmetrical: the sender stays lightweight to run on edge hardware without GPU acceleration, while the heavy reconstruction is run on the server side. The primary advantage of this design is that it allows MSIC to achieve compression gains from less important regions of the image, while maintaining high quality for more important regions regardless of compression levels. The framework is evaluated on the Cityscapes dataset against standard JPEG and SQ-GAN, a recent learned semantic compression baseline. At matched bitrate, the proposed method leads on structural similarity and preserves vehicles, road markings and other regions of interest at full quality. Therefore, the tradeoff favors MSIC in the type of content that downstream surveillance applications rely heavily upon.