In this paper, we propose a strategy for network simplification and acceleration. First, we propose to generate a suitable resized image using multiscale patching in the first convolutional layer, which can then be used for the rest of the network. We use p convolutional filters that operate on patches of size m × n, and we first select all the possible non-superposed m × n patches from the available images. If the number of such patches is not sufficient, the remaining ones are collected using scales a × b or c × d such that ab = cd = mn or a = 2 m or b = 2n. Patches generated from the former condition are directly extracted from the images, while the downsampled results are used in the latter case. We also introduce a 2-dimensional decomposition for patch compression, by stacking all the available image patches along the columns and applying a 2D PCA decomposition. Finally, a layer weight decomposition technique followed by module-based finetuning is adopted, for a new fast module-based CNN model. Extensive evaluations using public data sets like MNIST, Pascal VOC, and WebCASIA, and with different state-of-the art CNN architectures like Open-Face, VGG and DarkNet verify that our proposed model is able to accelerate the training process and even to provide higher classification accuracies for small-sized datasets. We obtain 8% increase in Top-1 and Top-5 recognition rates, and 5% increase in F1 score over general interpolation-based resizing.
