Efficient Methods Of Object Detection Based On Deep Neural Networks

Object detection is an important technology in the field of computer vision which is used for recognizing and locating all the interested objects appear in specific images.It is the main prior task for visual perception applications like automatic drive,video security and augmented reality,so its accuracy and efficiency will influence on the reliability and usability of many computer vision tasks.That's why a large number of universities,institutes and technology companies are focusing on it for a long time.Nowadays,object detection algorithms based on deep neural networks have improved the detection accuracy hugely,but their complex structures and considerable numbers of parameters make it difficult to implement them into use on mobile devices and real-time jobs.Although there are some optimization approaches in speed for this scenario,most of them will have harmful effects on performance.Against such backdrop,in this paper we investigate the existing object detection methods as well as network designing and optimization methods for deep neural networks,then propose an improved solution,in which one-stage object detection architecture is divided into two parts,feature extraction network and object detector network,and these two parts are optimized coordinately.At last,an accuracy and efficiency object detection algorithm is designed,and successfully used in a real-time outdoor sites detection platform on mobile devices.The detailed works in this paper include:For the problem that there are uncontrollable huge number of parameters in the existing feature extraction networks used in object detection algorithms based on deep neural networks,we design a new small-scale feature extraction network and apply a further optimization on it.Then considering the problem that the small-scale networks need to keep their accuracy while decreasing the number of layers and parameters,each parameter needs to provide a higher contribution to the result.We use the global consideration effect of attention models for parameter structure and relationship supervision,enhancing the quality of extracted features.For SSD architecture detectors,we change the feature extraction networks to light weighted ones.Then,in order to solve the problems that the low-level features which are not abstractive and semantic enough make small objects detection inaccurately,a fusion of low-level and separable deconvoluted high-level features is used for improve the quality of the features and detection accuracy.For objects with abnormal ratio,we research the CornerNet,a corner-based object detector,and transfer the ENet,which is a small-scale semantic segmentation network for feature extraction,proposing a fast and accurate modularization method of corner-based object detection.For the actual application scenario,we design and develop a real-time object detection platform on Android operating system based on the proposed methods.The related works are including:Data fetching and annotation,model designing and training as well as platform development.