HRPM Net: An Efficient Feature Learning Network from The Biological Modelling Of Human Retinal Perception Mechanism

Abstract

The biological model of the mammal visual mechanisms is very beneficial to feature learning in motionless images. It is proved that the visual mechanisms can improve the performance of the hand-crafted methods and CNNS method. Recently CNNs learn discriminate and robust features by changing the backbone, processing multi-scale feature maps, and adding attention mechanisms, etc. While they are relatively short of changing the network structure with human retina mechanisms, which have been proven to have a strong feature extract capability of images by traditional feature descriptors. To address this problem, we present two CNN blocks, multi-scale receptive field convolutional block (MSRF) and Sensitivity block (SENSI), both of which are constructed by modeling the human retina ganglion cell's mechanisms. MSRF is designed to enhance the feature discriminability and robustness by imitating the exponentially increased way of the receptive fields of the P ganglion cells in the human retina. We constructed experiments to get the specific value of the size of the receptive fields, and it can capture both local and global features with various convolution kernels. SENSI is presented to make sure each receptive field has a suitable weight to choose which receptive field can better learn features. Both of them help to learn features and can be easily integrated into the existing CNN models. The framework is evaluated on two benchmark datasets. We further assemble MSRF and SENSI to the top of SSD, constructing the HRPM Net. The model outperforms the state-of-the-art approaches by a considerable margin on MS COCO, VOC 2012, and VOC 2007 datasets. The results also show that MSRF block and SENSI block are helpful in feature learning and can improve the performance by a margin.