Dual Aggregation Convolution for Image Super-Resolution
Zhongxun Wang, Zheng Xie · 2024
In recent years, numerous image super-resolution (SR) methods have been introduced, achieving significant advancements. However, many existing methods employ complex, task-specific operations that, while effective, introduce substantial computational overhead when directly stacked, thereby hampering their practical utility. In response, we introduce DACNet, an efficient SR model that innovatively integrates spatial and channel information through a collaborative framework. Specifically, we alternately apply the Aggregation Spatial Convolution Module (ASCM) and the Aggregation Channel Convolution Module (ACCM) to dynamically extract and integrate multi-level features. Furthermore, we employ a Convolution Compensation Module (CCM) to mitigate local pixel forgetting. Our comprehensive experiments demonstrate that DACNet not only achieves superior performance but also maintains minimal computational demands, making it highly effective in resource-constrained environments.