Artificial neural network implementation on edge devices with limited resources is hindered by the computational bottleneck of image decompression and the inherent redundancy of visual data for machine-centric tasks. While compressed domain processing avoids inverse transforms, existing methods often process all frequency coefficients, ignoring task-specific redundancies. This paper proposes TAP-HybridNet, a unified framework that synergizes Task-Aware Compressed Domain Pro cessing (TAP-CDP) with the compact HybridConvNet architec ture. By integrating a learnable Task-Aware Coefficient Selection (TACS) module, the framework identifies and retains only the minimal spectral subset required for semantic inference. Exper imental results on CIFAR-10 demonstrate that TAP-HybridNet matches the accuracy of pixel-domain baselines within 0.5%, while reducing input dimensionality by approximately 60% and eliminating the costly Inverse Discrete Cosine Transform (IDCT) step. The framework establishes a superior Pareto frontier for Rate-Accuracy-Complexity (R-A-C) optimization in edge vision systems.

Index Terms: Compressed domain inference, DCT, Task aware compression, Hybrid architectures, Edge computing.