Lightweight Object Detection Based on Super-resolution

Abstract

Deep learning-based object detection has improved detection accuracy compared with traditional object detection. But there are problems such as large network parameters and high computer hardware requirements, making it difficult to meet the deployment on embedded or mobile devices. To address this, we optimize the classical SSD algorithm and propose a lightweight SSD object detection algorithm with super-resolution feature fusion. First, MobileNetv2 is used instead of VGG-16 as the backbone network. Secondly, five additional layers are added to generate feature maps of different sizes using the improved MixConv algorithm. Finally, the designed super-resolution feature fusion module generates a new feature pyramid. The test results on the PASCAL VOC dataset show that the model parameters are reduced by 76.26%, and the model complexity (FLOPs) is reduced by 83.13% compared to the original SSD algorithm. The detection speed increases by 1.27 times to 58.3 frames per second, and the average detection accuracy (mAP) can reach 72.7%. While ensuring accuracy, it reduces the requirement of the network on computer hardware and realizes the lightweight of the network.