A convolutional neural network model for object detection based on receptive field

Abstract

The mainstream methods for object detection can be divided into two types: one-stage (based on Integrated Convolutional Network) and two-stage (based on Candidate Box Convolutional Network). The one-stage method is fast but not accurate. While the two-stage method is accurate but slow. Thus, this paper proposes a novel convolutional neural network model that can satisfy both efficiency and accuracy needs for real-time object detection. Based on Single Shot Detector (SSD) and Feature Pyramid Networks (FPN), the proposed model addresses the issue of small object detection. The introduction of receptive field block (RFB) and RefineDet network improves the accuracy of the model. The experiment results show that the mAP value of the model exceeds 80%, and the FPS is above 30, when the size of the input image is 320 * 320.