Deep compressive joint source channel coding for efficient end-to-end communication systems

This thesis integrates compressive sensing (CS) and deep joint source-channel coding (JSCC) to design bandwidth-efficient, robust wireless image transmission systems optimized for downstream inference. Conventional models employ fixed-dimensional representations that fail to adapt to varying channels, wasting resources on pixel reconstruction when only task-level metrics are required. To resolve these limitations, this work develops a unified family of compressive, adaptive, and task-oriented JSCC frameworks. The first part introduces fixed-rate compressive JSCC schemes embedding learned sampling and two-stage reconstruction directly into end-to-end pipelines. Utilizing block-based sampling, generalized divisive normalization (GDN), and content-adaptive measurements, these architectures significantly improve reconstruction fidelity and robustness over deep JSCC baselines across additive white Gaussian noise (AWGN) and Rayleigh fading channels. The second part extends this to rate-adaptive operation, where the compression ratio dynamically adapts to both image content and channel signal-to-noise ratio (SNR). Within a single model, this framework supports multiple effective rates and delivers superior reconstruction quality at lower average bandwidths than fixed-rate or purely SNR-adaptive alternatives. The final part shifts from reconstruction to semantics, introducing task-oriented JSCC frameworks for applications like detection and recognition that transmit compressed feature embeddings instead of full images. Under severe channel noise, these task-aware architectures maintain high inference accuracy while drastically reducing channel uses, establishing a robust foundation for adaptive, task-aware wireless image transmission.