DSFNet: Class-Agnostic Object Counting Network With Dual Enhancement of Similarity and Feature Representation

Recently, the class-agnostic counting frameworks have gained popularity for general object counting, focusing on similarity matching between exemplar objects and query images. However, the existing methods ignore the noise interference in similarity matching and the feature representation of the direction variation of the target object, resulting in limited counting performance. In this article, we propose a novel class-agnostic counting network with dual enhancement for similarity and feature representation. Specifically, we propose a similarity enhancement aggregation module consisting of pooling mechanism enhancement and weighted feature aggregation to improve similarity correlation. To strengthen the feature representation of variable orientation angles and scales in the target object, we design a learnable directional feature enhancement module, which enriches the semantic information and improves counting in dense scenes. Notably, our method can also achieve competitive performance with limited training epochs in the zero-shot setting. Extensive experimental results show that our method achieves superior performance on the public benchmark FSC-147 and the vehicle dataset CARPK.