Enhanced Convolutional Neural Network Architecture with Attention-Guided Feature Refinement for Robust Image Classification

Abstract

Image classification remains a foundational problem in computer vision, underpinning applications that range from medical diagnostics to autonomous navigation. Although deep convolutional neural networks (CNNs) have substantially advanced classification accuracy over the last decade, standard architectures such as VGG, ResNet, and plain CNNs often struggle to jointly optimize accuracy, parameter efficiency, and robustness to intra-class variation. This paper proposes an AttentionGuided Refinement Network (AGR-Net), a CNN architecture that integrates a lightweight channel-and-spatial attention module with residual feature refinement blocks to improve discriminative feature learning while maintaining a compact parameter footprint. The proposed model is evaluated on the CIFAR-10 benchmark dataset using a rigorous experimental protocol that includes data augmentation, learning-rate scheduling, and five-fold cross-validation. Comparative experiments against baseline architectures, including a plain CNN, VGG-16, ResNet-18, and MobileNetV2, demonstrate that AGR-Net achieves a top-1 accuracy of 94.6%, outperforming the strongest baseline by 1.8 percentage points while using 27% fewer parameters than ResNet-18. Ablation studies confirm that the attention module and refinement blocks each contribute measurably to the accuracy gain. The results indicate that attention-guided refinement offers a favorable trade-off between accuracy, efficiency, and robustness, making the proposed architecture suitable for deployment in resource-constrained and real-time image classification systems. The paper further discusses the advantages, practical applications, limitations, and future research directions of the proposed approach.