WebAR Object Detection Method Based on Lightweight Multiscale Feature Fusion

Abstract

In the field of mobile augmented reality, object detection methods based on deep learning have become a research hotspot. Aiming at the problem of low detection accuracy and inability to meet the real-time detection target of mobile devices in the case of complex scenes and target occlusion, a WebAR object detection scheme with a lightweight neural network structure is proposed. First, using the lightweight network Mobilenet as the backbone feature extraction network of the SSD model of the object detection framework can significantly reduce the number of model parameters and computation and meet the real-time detection of targets. Second, a multi-scale feature fusion module is proposed. The Kronecker convolution and feature pyramid network are introduced to expand the receptive field. Capture richer multi-scale feature information, reduce the loss of important information, and improve the precision measurement accuracy. Finally, an object detection system based on WebAR is designed to realize the recognition of 2D images. The WebAR detection system designed with this model improves the efficiency of users' work. Experiments show that the improved model in this paper achieves better detection results on the PASCAL VOC2007 and MS COCO test dataset.