Pruning of Network Filters for Small Dataset

Abstract

Many existing object detection networks achieve outstanding accuracy in some open-source datasets, which usually contain a large number of images and target categories. However, a majority of the state-of-the-art networks use hundreds of filters and layers to extract plenty of features on behalf of getting a better result. Also, it is known that the networks tend to be redundant with so many parameters. In order to explain the redundancy more intuitively, we propose to visualize the feature maps, among which some is even found to impact negatively on the detection accuracy, especially when it comes to some small datasets containing only a few target categories. To address this problem, we propose the pruning rules of filters, whereby finding the useless filters and remaining the helpful ones based on feature maps, hence the networks are turned into thinner ones. Our final networks use filters less than one-tenth of the baseline model in some layers, but still outperform some related object detection methods.

Keywords

• Pruning strategy
• feature map
• convolution network
• small dataset