A Dynamic Feature Caching Mechanism for Accelerating CNN Real-time Inference

To address the computational efficiency bottle-neck faced by large-scale convolutional neural networks when deployed on edge devices, this paper proposes a dynamic feature caching mechanism for accelerating CNN inference. The mechanism enables intelligent reuse of convolutional features by establishing a similarity discrimination model between input data and historical features, thereby circumventing redundant computations. As a lightweight software strategy, it integrates into existing inference workflows without modifying model architectures or requiring specialized hardware. For continuous video frame processing scenarios, this mechanism achieves an $87 \%$ feature reuse rate while improving inference speed by $32.9 \%$ and maintaining high accuracy. The solution provides low-intrusive acceleration support for resource-sensitive environments such as mobile terminals, significantly reducing computational loads on edge devices.