Real-Time Object Detection Based On Multi-Scale Feature Fusion

The precise positioning and object recognition in videos and images play an important role in many practical scenarios,such as intelligent video surveillance,content-based image retrieval and navigation etc.At present,the target detection algorithm based on deep learning made a breakthrough,the detection accuracy compared with the traditional method has reached a new height.However,in some applications with high real-time requirements,the existing real-time detection network is hard to achieve high detection accuracy.Therefore,how to ensure real-time detection and achieve precise location and recognition of objects is still an urgent problem in this field.On the basis of the real-time detection network,SSD,by improving the feature extraction and target prediction method,we propose a real-time target detection network based on multi-scale feature fusion,thus improving the positioning accuracy and recognition of the scale accuracy of target.Specific improvements include the following two aspects:First,aiming at the location accuracy problem of object prediction layer in SSD network due to the low resolution,in this paper,we analysis the multi-scale features extracted by SSD network.Then the position sensitive information provided by the lower layer's detail features and the context information provided by the high-level semantic features are combined through feature fusion method.And different feature extraction and fusion methods are discussed.The positioning accuracy of the target prediction layer in SSD network is effectively improved.The experimental results show that the feature fusion SSD network gets some improvement in the accuracy of localization compared to the original network.Secondly,because the objects in SSD network prediction layer(especially shallow layer target prediction)lack enough semantic information,leading to the object recognition accuracy is not high.This paper introduces a design of a feature embedded(Inception)prediction structure.So that the object features' semantic of the SSD prediction layer is strengthened without changing the spatial resolution,and low scale characteristics are also embedded in the high semantic features to achieve collaborative prediction,so as to improve the object recognition accuracy of SSD network on each scale.The experimental results show that by combining the above two aspects of improvement,the proposed feature fusion and inception prediction network gets a great improvement in position accuracy and the recognition accuracy than the original SSD.The detection accuracy is improved from 77.4% to 79.7% in the PASCAL VOC2007 test data set.Showing a great advantage compared with present excellent target detection algorithms.