| Title |
Intrinsic Dimensionality based Layer-wise Channel Pruning for Object Detection Network Slimming |
| Authors |
이연주(Youn Joo Lee) ; 서재규(Jae Kyu Suhr) ; 정호기(Ho Gi Jung) |
| DOI |
https://doi.org/10.5573/ieie.2023.60.3.43 |
| Keywords |
Intrinsic dimensionality; Layer-wise channel pruning; Network slimming; Object detection; YOLOv4 |
| Abstract |
Recently, network slimming is actively researched to mount CNNs that provide high accuracy in resource-limited embedded systems such as autonomous vehicles, smartphones, and drones. Among the network slimming methods, the channel pruning method is widely used, which removes unimportant channels or filters in the network. Previous channel pruning methods use an only one threshold through all layers to check channel’s importance. CNNs have a hierarchical structure in which multiple layers are cascaded. Based on this point, this paper proposes a new layer-wise channel pruning method based on intrinsic dimensionality which can be used to estimate the number of channels required for each layer. In the experiment, the proposed method was applied to YOLOv4 proven superior in terms of accuracy and inference time and was evaluated on PASCAL VOC dataset and VisDrone dataset. Experimental results showed that the proposed method has better mAP performance on VOC and VisDrone (except 70% pruning ratio) datasets than the previous method using a global threshold. |