Modified Selective Kernel Networks and Application in Classification

In standard convolutional neural network, the kernel size or receptive field in each layer is same. It is not the same case with neurons in the primary visual cortex of humans which can be modulated by stimulus. The size of receptive field plays a very important role in recognizing multi-scale objects. In this thesis, our aim is to propose a method which has the ability to change the receptive field size in neurons of the same layer. This is enable the CNNs to detect multi-scale objects efficiently. In the paper, we have described various state of the art architectures which can be used as backbones for our method. We discuss how our model can be used for image search engines ranking.