DLNet: Domain-specific Lightweight Network for On-Device Object Detection

A deep neural network (DNN) is one of the key technique in many intelligent applications. Since DNN models have many parameters, they require machines to perform computation-intensive tasks. Therefore, it is difficult to apply the DNN model to resource-constrained environments. In this paper, we propose the domain-specific lightweight network (DLNet) for on-device object detection. Every domain has the classes group that contains objects with a high frequency of appearance in domains. Since the DLNet is trained to detect the classes group of each domains, it becomes more shallower and narrower than existing detection models. Therefore, we can reduce the number of parameters and runtime of object detection. To evaluate the performance of the proposed model, we conduct comparative experiments with YOLOv3 and Tiny-YOLO. The precision and recall are 9.9 % and 14 % higher than those of YOLOv3 and Tiny-YOLO, respectively. The results shows that the DLNet achieve higher efficiency and better performance on non-GPU devices.