Lightweight the Focus module in YOLOv5 by Dilated Convolution

Authors:

Abstract
Focus is an advanced module in YOLOv5. However, no detailed work has been published to demonstrate its effectiveness so far. In this paper, we analyze the Focus module from two respects, FLOPs and receptive field, and reveal that it’s approximated to convolution layer with the same kernel size, twice the receptive field and quadruple the computational cost. Inspired by dilated convolution, we propose a new module named DFocus. DFocus is comparable in performance to Focus but without any additional computation cost. The experiments on PASCAL VOC dataset demonstrate the effectiveness of our work.

This publication has 13 references indexed in Scilit: