Design of an Efficient CNN Architecture with SGD Optimization for Accurate and Robust Object Classification

Abstract

different models, Convolutional Neural Networks (CNNs) perform better than traditional image processing methods and simple neural networks. Despite their success, achieving a balance between classification accuracy, model complexity, and computational efficiency remains a key research challenge as classification accuracy and training cost strongly depend on network structure and learning strategy. In this study, an optimized deep CNN model for object classification is explored by carefully analyzing important architectural and training hyperparameters. Experiments are conducted using CIFAR-10 standard datasets. The proposed architecture is organized into three convolutional stages with increasing channel capacity. Batch normalization is utilized to stabilize gradient propagation, while max pooling is employed for spatial compression. Dropout-based regularization is strategically integrated to reduce overfitting. Additionally, adaptive average pooling is used before the classifier to maintain expressive feature learning while minimizing the overall parameter count. All experiments are conducted using GPU acceleration and optimized using momentum based minibatch Stochastic Gradient Descent (MSGD) algorithm to ensure stable and efficient learning. Model performance is evaluated in terms of classification accuracy, parameter efficiency, training time and number of training epochs. The experimental results demonstrate that the proposed CNN architecture achieves competitive accuracy on CIFAR-10 while maintaining a relatively low computational cost, highlighting the effectiveness of careful architectural design and optimization for efficient image classification. In addition, comparisons with existing state-of-the-art models show that the proposed method delivers higher classification accuracy and improved computational efficiency.