Accurate Counting Network with Multi-Scale Denoising

Object counting plays a crucial role in production and daily life. Although prior studies have achieved encouraging results, current methods still face challenges in complex scenarios such as uneven distribution of image targets, dense tiny targets, and changes in picture viewpoints. Under these conditions, features extracted by backbone networks contain rich information but exhibit redundancy. While redundant information improves overall feature diversity, it concurrently introduces substantial noise interference. To address these issues, we propose a Feature Enhancement Fusion Module (FEFM) and a Noise-Resistant Feature Block (NRB). The NRB, with its simple architecture, effectively enhances effective information while suppressing noise. The FEFM integrates multiple NRB-based convolutional units to achieve multi-scale feature enhancement and fusion, significantly improving counting accuracy. By incorporating NRB and FEFM into existing counting models, we substantially boost their performance. We evaluate the performance of the module on three counting public datasets: the crowd counting benchmark ShangHaiTechPartA, ShangHaiTechPartB dataset, and the object counting benchmark CARPK dataset. Competitive results on the ShangHaiTechPartB crowd counting dataset and new state-of-the-art performance on the ShangHaiTechPartA, CARPK dataset demonstrate the effectiveness and robustness of FEFM with NRB.