Attention-Enhanced Deep Convolutional Networks for Multi-Scale Feature Learning in Complex Image Classification

Abstract

Deep convolutional neural networks (CNNs) have achieved remarkable success in image classification; however, their ability to effectively capture multi-scale features in complex visual environments remains a challenge. This study proposes an attention-enhanced deep convolutional network designed to improve multi-scale feature learning for complex image classification tasks. The framework integrates attention mechanisms with hierarchical convolutional structures to dynamically emphasize informative regions while suppressing irrelevant features. The proposed model combines spatial and channel attention modules with multi-scale feature extraction layers, enabling improved representation of both fine-grained and global contextual information. Experimental evaluation on benchmark image datasets demonstrates that the attention-enhanced architecture achieves superior classification accuracy compared to conventional CNN models, particularly in scenarios involving cluttered backgrounds and high intra-class variability. The results also show improved convergence behavior and robustness to noise and scale variations. Furthermore, the study investigates optimization strategies such as feature fusion, residual learning, and adaptive pooling to enhance performance. The findings indicate that attention mechanisms significantly improve feature discrimination and model interpretability. This research contributes a scalable and efficient architecture for advanced image classification tasks, with applications in medical imaging, remote sensing, and intelligent vision systems.