| Title |
An Importance Search Filter Pruning Method for Accelerating Object Detection Network in Embedded Systems |
| DOI |
https://doi.org/10.5573/ieie.2024.61.5.31 |
| Keywords |
Object detection; Network compression; Pruning; Inference time |
| Abstract |
In recent years, with the development of computer technology, research on CNN-based object detection networks has been actively conducted. However, a large number of CNNs can make inference difficult in embedded environments with limited memory and computation. A typical solution to this problem is network pruning. Network pruning can facilitate inference on embedded boards by reducing the amount of memory and computation required by removing redundant parameters. However, most pruning methods require two stages of training, which consumes a lot of time and resources, and cannot guarantee an optimal lightweight network because they cannot reflect the changes in channel relationships due to pruning. Therefore, this paper proposes an importance search method to obtain an optimal lightweight network, and simplifies the pruning process to propose an importance search filter pruning method that can be pruned with only one stage of training. In this paper, we apply pruning to the SSD network with VGG-16 and ResNet-50 as the backbone network, and measure the inference speed on Jetson Xavier NX. In the network using ResNet-50, the experimental results showed that mAP(0.5) decreased by 0.5%, 0.7%, and 1.0% depending on the pruning ratio, but inference time improved by 12.75%, 16.03%, and 21.66%. In addition, the learning time is up to 43.85% faster than other methods and has high performance when compared to networks with similar pruning ratios. |