Adaptive pruning threshold based convolutional neural network for object detection

In this paper, it is proposed that the redundancy in convolutional neural networks of object detection can be effectively removed via an adaptive pruning threshold method (APTCNN) which is associated with scaling factors in batch normalization layers. In this way, the channel pruning can be done iteratively with varying pruning threshold until the satisfactory performance is obtained. The method is also useful for identifying the unimportant convolutional layers. Therefore it can be applied for layer pruning. The experiments are conducted on three benchmark object detection datasets. APTCNN is verified for pruning the backbone network of object detectors YOLOv3 and YOLOv3-spp. It is shown that the importance of channels and layers are accurately ranked by the proposed adaptive threshold. For the channel pruning, our method reduces the size of YOLOv3 and YOLOv3-spp by 32x and 48x respectively, and accelerates 1.7x and 1.9x respectively. However, the accuracy suffers only 0.77% and 1.32% loss, respectively. As a result, the redundancy in the network architecture can be efficiently removed yielding a slimmed model that has lower computing operations, reduced size, and without compromising accuracy.