DRFnet: Dynamic receptive field network for object detection and image recognition.

Abstract

Biological experiments discovered that the receptive field of neurons in the primary visual cortex of an animal's visual system is dynamic and capable of being altered by the sensory context. However, in a typical convolution neural network (CNN), a unit's response only comes from a fixed receptive field, which is generally determined by the preset kernel size in each layer. In this work, we simulate the dynamic receptive field mechanism in the biological visual system (BVS) for application in object detection and image recognition. We proposed a Dynamic Receptive Field module (DRF), which can realize the global information-guided responses under the premise of a slight increase in parameters and computational cost. Specifically, we design a transformer-style DRF module, which defines the correlation coefficient between two feature points by their relative distance. For an input feature map, we first divide the relative distance corresponding to different receptive field regions between the target feature point and its surrounding feature points into N different discrete levels. Then, a vector containing N different weights is automatically learned from the dataset and assigned to each feature point, according to the calculated discrete level that this feature point belongs. In this way, we achieve a correlation matrix primarily measuring the relationship between the target feature point and its surrounding feature points. The DRF-processed responses of each feature point are computed by multiplying its corresponding correlation matrix with the input feature map, which computationally equals to accomplish a weighted sum of all feature points exploiting the global and long-range information as the weight. Finally, by superimposing the local responses calculated by a traditional convolution layer with DRF responses, our proposed approach can integrate the rich context information among neighbors and the long-range dependencies of background into the feature maps. With the proposed DRF module, we achieved significant performance improvement on four benchmark datasets for both tasks of object detection and image recognition. Furthermore, we also proposed a new matching strategy that can improve the detection results of small targets compared with the traditional IOU-max matching strategy.