A Two-Path Network for Cell Counting

The density map is an effective manner to infer how many cells a cell image contains, and it carries valuable information. However, a fine-grained density map requires rich spatial information to recover the distribution details. In this paper, we propose a cell counting network with two paths, i.e. detail path and context path, which respectively extract spatial details and semantics. The detail path encodes the spatial information with small convolutional kernels. The context path rapidly enlarges the receptive field and extracts multi-scale features with an atrous spatial pyramid pooling. At the end of the two paths, we design a feature fusion module to merge the high-level feature maps from the two paths. To decrease the parameters and computation, we directly upsample the fused feature maps to the input size and decode them to obtain the density map. The proposed model is evaluated on three cell datasets and a popular crowd dataset Shanghaitech Part-B. The experiments illustrated that the proposed model not only achieves superior performance on cell datasets but also generalizes well on the crowd dataset.

Abstract licence: CC BY-NC-ND

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.