Less Is More: Matched Wavelet Pooling-Based Light-Weight CNNs With Application to Image Classification

We are concerned with the development of effective training strategies for light-weight deep networks, which facilitates harnessing power learning on hand-held devices. A vast literature has focused enhancing performance convolution models (CNNs) by introducing changes in model architecture and/or developing new procedures. Recent witnessed a growing focus pooling layers, one example this is wavelet pooling. Pooling often adopted CNNs to enlarge receptive field, yet traditional methods cause information loss undermines further operations such as feature extraction. By capitalizing spectral information, (WP) addressed drawback neighborhood pooling, and we have demonstrated its positive impact CNN architectures MobileNets previous research. In paper, report following significant observation: Including all sub-bands decomposition does not necessarily yield higher accuracies than using certain sub- set sub-bands, i.e., less more. identifying be included process every image during testing, propose inference procedures, namely, Matched Wavelet (MWP), architectures, particular, MobileNets. As an example, classification application considered. On two widely datasets, proposed MWP algorithm consistently outperform attained average 20% accuracy increase compared recent literature.