PROJECT TITLE :

Learning Versatile Convolution Filters for Efficient Visual Recognition

ABSTRACT:

This article presents versatile filters that can be used to construct efficient convolutional neural networks, which are widely used in a variety of visual recognition tasks. In order to meet the requirements of effective deep learning techniques that can be implemented on reasonably priced hardware, a number of different approaches to learning compact neural networks have been developed. The majority of these studies investigate small, sparse, or quantized filters as potential ways to reduce the size of existing filtering systems. On the other hand, we approach the problem of filters from an additive standpoint. With the assistance of binary masks, a primary filter can give rise to a number of secondary filters that can be further refined. These secondary filters all inherit in the primary filter without occupying any additional storage space; however, once they have been unfolded in computation, they have the potential to significantly improve the filter's capability by integrating information that has been extracted from a variety of receptive fields. In addition to the investigation of spatial versatile filters, we also look into versatile filters from the channel's point of view. It is possible to further customize binary masks for a variety of primary filters while adhering to orthogonal constraints. We perform theoretical analysis on the complexity of networks, and we introduce a convolution scheme that is both effective and efficient. The results of our experiments on benchmark datasets and neural networks indicate that our adaptable filters are capable of achieving an accuracy that is comparable to that of the original filters, but they require less memory and have lower computation costs.