Gradually Global Pruning for Object Detection Network

State-of-the-art object detection methods employ the deep convolutional neural Network (CNN) to achieve excellent results on several public datasets. However, most of these methods depend on complex backbone networks which limit the application in many real-world scenarios. In this paper, a novel gradual global pruning is presented to remove the redundant filters for object detection models, which can largely reduce the model size of the network while maintain the network accuracy. Firstly, the importance of individual filter is evaluated globally across all the network layers by calculating the sum of its absolute weights. Then, a progressive sparsity control function is proposed to gradually reduce the proportion of pruning in the implementation process. Finally, this paper finetunes the pruned network to restore its accuracy. The experimental results on Pascal VOC2007 dataset show that the proposed method can drastically reduce the model size of the network with negligible loss in detection accuracy.