Rega-Net: Retina Gabor Attention for Deep Convolutional Neural Networks

Extensive research works demonstrate that the attention mechanism in convolutional neural networks (CNNs) effectively improves accuracy. Nevertheless, few design mechanisms using large receptive fields. In this work, we propose a novel method named Rega-Net to increase CNN accuracy by enlarging field. To best of our knowledge, increasing field network requires size convolution kernel, which also increases number parameters. For solving problem, kernels resemble non-uniformly distributed structure inspired human retina. Then, sample variable-resolution values Gabor function distribution and fill these retina-like kernels. This allows essential features be more visible center position We further an module including Experiments achieves 79.96% Top-1 on ImageNet-1K for classification 43.1% mAP COCO2017 object detection. The increased up 3.5% compared baseline networks.