Dynamic Global Query Fusion: A Plug‐and‐Play Module for Enhancing Convolutional Networks

ABSTRACT In recent years, self‐attention mechanisms have demonstrated remarkable performance across various computer vision tasks, gradually emerging as a mainstream approach. However, compared to traditional Convolutional Neural Networks (CNNs), its high quadratic complexity and limited adaptability to 2D structures have constrained its broader application and adoption. To enhance the feature extraction capability of CNNs, this work focuses on augmenting input information and introduces a novel architectural unit called the Global Query Vector (GQ Vector). The proposed unit adopts a co‐evolutionary architecture consisting of a parallel branch and the main backbone network, which continuously integrates and refines global semantic information during forward propagation, establishing a cross‐layer, persistent context memory mechanism. This design enables progressive accumulation and refinement of contextual information, thereby enhancing the CNN's capacity to model long‐range dependencies. Building on this, we propose a novel CNN architecture named a Global Query Convolutional Network (GQConvNet). It can be seamlessly integrated into existing CNN frameworks, further enhancing their performance. For example, on the ImageNet‐1K dataset, a ResNet‐50 model augmented with GQ Vector achieves a 1.7% improvement in Top‐1 accuracy over baseline models. This work offers a fresh perspective on optimizing CNNs, with substantial academic value and practical implications.