Infrared and Laser Engineering, Volume. 50, Issue 8, 20200510(2021)
Infrared object detection network compression using Lp normalized weight
Figures & Tables(9)
Fig. 2. Sparsity of weight with respect to
Fig. 3. Training process of sparse neural network for object detection
Fig. 4. Result comparison of infrared object detection between classical neural networks and sparse neural networks
Table 1. Simulated infrared dataset
View tableView in ArticleTable 1. Simulated infrared dataset
Classification Training Test Total Class 1 208 28 236 Class 2 210 26 236 Class 3 219 30 249 Class 4 192 29 221 Total 829 113 942
Table 2. Object detection model and result on simulated infrared dataset
View tableView in ArticleTable 2. Object detection model and result on simulated infrared dataset
Method Scale AP mAP Backbone Detector Class 1 Class 2 Class 3 Class 4 Faster R-CNN Dense 26 852 416 14 511 140 0.912 0.885 0.927 0.972 0.925 Sparse 5 337 352 14 511 130 0.910 0.875 0.936 0.982 0.926 SSD300 Dense 22 943 936 1 202 958 0.893 0.879 0.914 0.965 0.914 Sparse 4 103 396 1 202 958 0.889 0.867 0.924 0.981 0.917 YOLOv3 Dense 55 294 688 6 245 196 0.914 0.898 0.919 0.972 0.926 Sparse 14 829 742 6 245 196 0.906 0.895 0.927 0.984 0.928
Table 3. Object detection model and result on VOC2007 dataset
View tableView in ArticleTable 3. Object detection model and result on VOC2007 dataset
Method Faster R-CNN SSD 300 YOLOv3 Dense Sparse Dense Sparse Dense Sparse Nonzero parameters Backbone 26 852 416 15 756 216 22 943 936 14 995 952 55 294 688 37 291 638 Detector 14 593 140 14 593 140 3 341 550 3 341 550 6 331 357 6 331 357 AP Aero 0.833 0.826 0.854 0.847 0.801 0.802 Bike 0.781 0.773 0.798 0.795 0.848 0.845 Bird 0.735 0.737 0.702 0.712 0.716 0.726 Boat 0.532 0.528 0.568 0.543 0.652 0.641 Bottle 0.487 0.493 0.457 0.474 0.638 0.647 Bus 0.774 0.765 0.790 0.781 0.861 0.858 Car 0.745 0.748 0.757 0.752 0.858 0.859 Cat 0.887 0.872 0.756 0.765 0.847 0.857 Chair 0.449 0.443 0.871 0.865 0.547 0.541 Cow 0.765 0.771 0.524 0.542 0.715 0.725 Table 0.548 0.536 0.768 0.764 0.690 0.681 Dog 0.865 0.857 0.605 0.612 0.828 0.827 Horse 0.817 0.825 0.868 0.874 0.842 0.846 Mbike 0.804 0.798 0.824 0.846 0.821 0.831 Person 0.794 0.782 0.820 0.811 0.807 0.802 Plant 0.391 0.387 0.458 0.447 0.441 0.437 Sheep 0.723 0.725 0.752 0.747 0.696 0.688 Sofa 0.608 0.595 0.691 0.698 0.699 0.696 Train 0.809 0.814 0.809 0.812 0.825 0.834 Tv 0.612 0.607 0.672 0.667 0.718 0.722 mAP 0.698 0.694 0.717 0.718 0.742 0.743
Tools
Get Citation
Copy Citation Text
Weipeng Li, Xiaogang Yang, Chuanxiang Li, Ruitao Lu, Xueli Xie, Chuan He. Infrared object detection network compression using Lp normalized weight[J]. Infrared and Laser Engineering, 2021, 50(8): 20200510
Paper Information
Category: Infrared technology and application
Received: Apr. 10, 2021
Accepted: --
Published Online: Nov. 2, 2021
The Author Email:
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20