| Abstract: |
The rapid proliferation of edge Artificial Intelligence (AI) systems has accelerated the deployment of image processing pipelines in resource-constrained environments such as Internet of Things (IoT) nodes, smart cameras, embedded processors, and autonomous systems. These deployments present two intertwined and critical challenges: ensuring energy efficiency under limited computational budgets and guaranteeing privacy preservation when sensitive visual data is processed at or near the point of capture. This review paper presents a comprehensive meta-analysis of prior research addressing these dual imperatives in edge AI image processing. Through a systematic survey of literature published between 2015 and 2024, this work synthesises findings from over one hundred studies spanning model compression, hardware-software co-design, federated learning, differential privacy, homomorphic encryption, and on-device inference optimisation. The review identifies dominant trends including the growing adoption of lightweight convolutional neural networks (CNNs), model quantisation, neural architecture search (NAS), secure multi-party computation, and privacy-by-design frameworks for vision tasks. Critical gaps are identified, including the absence of unified benchmarks that simultaneously evaluate energy cost and privacy loss, insufficient attention to adversarial robustness under energy constraints, and the limited applicability of current privacy mechanisms at sub-milliwatt inference regimes. The paper further discusses open research challenges and proposes a consolidated taxonomy to guide future work at the intersection of green edge AI and trustworthy visual computing. |
