Convolutional Neural Networks have been extensively used for solving many vision problems. However, due to high memory and computational requirements, deployment of these models on edge devices is limited. Many embedded friendly models such as MobileNet, ShuffleNet, SqueezeNet, and many more are proposed to serve this purpose. But these models are still not compact enough to deploy on edge devices. The popular metric-based pruning methods (which are aimed at pruning insignificant and redundant filters) could achieve limited compression for embedded friendly models such as MobileNet. In this paper, we propose a novel hybrid filter pruning method that prunes both redundant and insignificant filters at the same time. Additionally, we have designed custom regularizers that enable us to prune additional filters from convolutional layers. Pruning experiments are conducted on MobileNetv1 based Single-Shot Object Detector (SSD) for face detection problem. Through our experiments, we could prune 40.11% of parameters and reduce 67.03% of FLOPs from MobileNetv1 with a little drop in model performance (1.67 mAP on MS COCO). On an ARM-based edge device, the inference time is reduced from 198ms to 84ms.

Directionlets allow a construction of perfect reconstruction and critically sampled multidirectional anisotropic basis, yet retaining the separable filtering of standard wavelet transform. However, due to the spatially varying filtering and downsampling direction, it is forced to apply spatial segmentation and process each segment independently. Because of this independent processing of the image segments, directionlets suffer from the following two major limitations when applied to, say, image coding. First, failure to exploit the correlation across block boundaries degrades the coding performance and also induces blocking artifacts, thus making it mandatory to use deblocking filter at low bit rates. Second, spatial scalability, i.e., minimum segment size or the number of levels of the transform, is limited due to independent processing of segments. We show that, with simple modifications in the block boundaries, we can overcome these limitations by, what we call, in-phase lifting implementation of directionlets. In the context of directionlets using in-phase lifting, we identify different possible groups of downsampling matrices that would allow the construction of a multilevel transform without forcing independent processing of segments both with and without any modifications in the segment boundary. Experimental results in image coding show objective and subjective improvements when compared with the directionlets applied independently on each image segment. As an application, using both the in-phase lifting implementation of directionlets and the adaptive directional lifting, we have constructed an adaptive directional wavelet transform, which has shown improved image coding performance over these adaptive directional wavelet transforms.

In this paper we consider the problem of finding the direction of the content in a given image segment to foster the development of adaptive directional decompositions of images. We propose a mechanism based on a new directional metric that we define as Directional Variance (DirVar), a measure of variance along the discrete lines in the given direction. The simulation results in section III, within the frame work of directionlets, and in the lifting based directional DWT, show that our mechanism is very accurate and reliable in finding the content direction and saves lot of computations.