Attention-Guided Dynamic Re-Parameterization for Enhancing Multi-Scale Feature Fusion in Convolutional Neural Networks

Deep learning has become an indispensable tool in computer vision, enabling unprecedented performance in tasks such as classification, object detection and semantic segmentation. As advances in imaging technologies drive growth in the volume and complexity of visual data, there is an increasing demand for lightweight yet powerful neural network architectures capable of capturing rich multi-scale features efficiently. Structural re-parameterization has emerged as a promising strategy to optimize convolutional neural networks (CNNs) by simplifying inference while retaining expressive capacity. However, existing approaches (e.g., RepVGG, ACNet, DBB) often perform only static kernel fusion during inference, ignoring the semantic diversity and unequal importance of features captured across branches during training. This leads to partial information loss and constrained generalization. To address this challenge, we propose Attention-Guided Structural Reparameterization (AGSR), a novel and fundamental deep learning block that integrates multi-scale and multi-complexity convolutional branches with embedded attention mechanisms. By adaptively adjusting feature weights within each branch, AGSR enhances feature ex-traction and preserves critical semantic information throughout training. During inference, AGSR is seamlessly converted into a compact two-branch convolutional structure, maintaining efficiency without compromising performance. Experimental results on diverse benchmark datasets highlight the AGSR Block's remarkable performance enhancements. On the MVTec-D2S dataset, AGSR-enhanced versions of DenseNet, ResNet and MobileNet models achieved accuracies of 96.7%, 87.4% and 94.6%, respectively, surpassing baseline models and prior methods such as ACNet and DBB. Additionally, on the UC Merced Land Use dataset, AGSR elevated DenseNet's accuracy from 76.6% to 84.8%. These results affirm AGSR’s effectiveness in multi-scale feature fusion and its substantial potential for image processing.