Efficient and Robust Convolutional Neural Network Design for Resource-Constrained Image Recognition Systems

Convolutional Neural Networks (CNNs) have achieved state-of-the-art performance in visual recognition tasks through hierarchical feature learning . However, recent architectures increasingly rely on aggressive depth and parameter scaling, resulting in substantial computational overhead that restricts deployment on resource-constrained platforms. This paper introduces LiteRobustNet, a lightweight yet robust CNN framework jointly optimized for classification accuracy, computational efficiency, and real-world reliability. Three architectures—a baseline CNN, a deep CNN, and the proposed optimized model—are systematically evaluated on the CIFAR-10 dataset across predictive performance, parameter efficiency, inference latency, robustness to common image distortions, and deployment-level optimization. Experimental results demonstrate that LiteRobustNet reduces parameter count by 95.8\%relative to the deep CNN while retaining nearly 90\% of its classification accuracy. Robustness-aware training significantly improves stability under noise, blur, and resolution degradation \cite{6}, \cite{7}. Furthermore, post-training quantization compresses the model to 0.066 MB and accelerates inference by approximately 40\%, enabling practical edge deployment. These findings highlight the effectiveness of efficiency-aware architectural design combined with robustness optimization for real-time intelligent vision systems..