ANALYSIS OF CONVOLUTIONAL NEURAL NETWORK COMPRESSION METHODS FOR EFFECTIVE DEPLOYMENT IN EDGE AI ENVIRONMENTS

Abstract

The article is devoted to the research and empirical evaluation of convolutional neural network compression methods for their effective deployment in the Edge AI environment. Despite their high accuracy, traditional CNN architectures, such as ResNet-18, are too resource-intensive for peripheral devices with limited computing power, RAM, and energy consumption. The main focus is on finding the optimal balance between reducing resource consumption and maintaining high classification accuracy. The goal of this work is to investigate and demonstrate the effectiveness of special model compression techniques, including quantization, pruning, and knowledge distillation, for successfully transferring the powerful capabilities of CNNs to edge devices. The scientific novelty lies in a comprehensive, quantitative comparison of the impact of three main optimization techniques on key model performance indicators. Demonstration that full integer quantization (PTQ Int8) provides a compression ratio of 11.06x with minimal accuracy loss (0.0030), confirming it as the optimal first step. A comparative analysis proving that unstructured compression (50% of ResNet-18 weights) fully recovers and exceeds the baseline accuracy after fine-tuning, while structured compression leads to irreversible accuracy loss (up to 45.70%) under limited retraining conditions, requiring a more balanced approach. Confirmation that knowledge distillation allows the MobileNetV2 model to outperform its traditionally trained version (91.8% vs. 89.5%), maximizing accuracy under severe architectural constraints. Conclusion. Model compression is an engineering trade-off and a necessary condition for creating highly efficient, lowlatency, and energy-efficient deep learning solutions that can be successfully deployed in edge computing environments. The use of quantization allows energy-intensive models to be transformed into practical Edge AI solutions.