The Research On Lightweight Visual Object Detection Technology

With the vigorous construction of smart city,the volume of terminal devices such as surveillance cameras has skyrocketed,and the demand for intelligent analysis of terminal devices is also increasing.As an indispensable foundation in computer vision tasks,object detection has received extensive attention and research due to the important application value.In recent years,the rapid development of deep learning has greatly improved the performance of object detection.However,most current object detection networks cannot be used in the devices with limited computation power and memory resource such as electronic chips,mobile phones,etc.due to their large amount of calculations and parameters.To achieve a lightweight visual object detection network for resource constrained scenario,based on deep learning method,two lightweight visual object detection algorithms are proposed.(1)Using lightweight and efficient convolution techniques such as deep convolution,an innovative and efficient architecture is proposed,named 2-way Merging Lightweight Dense Block(2-way MLDB),The main innovation point is that2-way MLDB merges the duplicate parts of two independent branches in a dense block of the backbone network to obtain multi-receptive field features with fewer parameters and computations.Applying it to a carefully designed backbone network framework,we can obtain a lightweight model with higher detection accuracy.The best performing network structure is called as a lightweight object detection network based on branch merging.(2)Aiming at the problem of the decrease of accuracy caused by drastically reduced parameter size,an FPN-like SSD detection head based on an Attention Prediction Block(APB)structure is constructed,with a small increase in the amount of parameters and calculations,multi-scale feature fusion is achieved,and the most discriminative detection features at each scale are obtained,so that the network obtains a large improvement in detection accuracy,especially for small objects.The best performing network is called BMNet.Through extensive experiments on two classic benchmarks(PASCAL VOC and MS COCO),we demonstrate that the two algorithms proposed in this paper are superior to the most advanced lightweight object detection solutions such as Tiny SSD,Mobile Net-SSD,Mobile Netv2-SSD and Pelee in terms of parameter size,FLOPs and accuracy.Concretely,after fusing the two algorithms,the best model achieves 77.05% of m AP on PASCAL VOC 2007 test dataset with only 1.49 M parameters and 1.51 B FLOPs,whose resource requirements are relatively low and which can be well applied to resource limitation scenarios.