基于多尺度特征增强的小样本计数

目标物体计数是计算机视觉领域的重要研究方向。针对小样本计数中存在的样本与查询图像目标物体尺寸不一致、目标物体分布不均匀的问题，该文提出了多尺度特征增强计数算法。首先，基于特征金字塔构建自上而下的特征融合网络。在各级尺度上对查询图像中和样本相似度较高的区域进行样本特征增强，随后送入上一级特征匹配。然后，将各级增强后的查询特征送入回归器中，得到各级密度图。最后，求和，生成高质量的密度图。该文在FSC-147和CARPK数据集上进行测试。实验结果表明，该文所提模型的性能优于大多数其他方法，有效改善了目标物体大小变化造成的计数精度低的问题。

Multi-scale Feature Enhancement for Few-shot Object Counting

Object counting is an important research direction of computer vision. In order to solve the problem that the size of the object in the few-shot object counting is inconsistent with that in the query image, and the object distribution is uneven, a multi-scale feature enhancement counting algorithm is proposed. First, a top-down feature fusion network is constructed based on the feature pyramid. At each level of scale, the sample features of the regions with high sample similarity in the query image are enhanced, and then sent to the upper level for feature matching. The enhanced query features at all levels are sent into the regression head to obtain the density maps at all levels, and finally the sum can generate a high-quality density map. The experimental results on FSC-147 and CARPK datasets show that the performance of the model is better than that of most previous methods, which effectively improves the problem of low counting accuracy caused by the change of target size.