A Simple and Effective Convolutional Filter Pruning based on Filter Dissimilarity Analysis

In this paper, a simple and effective filter pruning method is proposed to simplify the deep convolutional neural network (CNN) and accelerate learning. The proposed method selects the important filters and discards the unimportant ones based on filter dissimilarity analysis. The proposed method searches for filters with decent representative ability and less redundancy, discarding the others. The representative ability and redundancy contained in the filter is evaluated by its correlation with currently selected filters and left over unselected filters. Moreover, the proposed method uses an iterative procedure, so that less representative filters can be discarded evenly from the entire model. The experimental analysis confirmed that a simple filter pruning method can reduce floating point operations (FLOPs) of TernausNet by up to 89.65% on an INRIA Aerial Image Labeling dataset with an only marginal drop in the original accuracy. Furthermore, the proposed method shows promising results in comparison with other state-of-the-art methods.