Deep Learning Based End-to-end Object Detection And Attributes Analysis Algorithm And Its Applications

With the dramatically increasing volume of multimedia data such as images and videos,to automatically and effectively analyze and understand the large amount of multimedia data by computers becomes a hot research topic in artificial intelligence.There are two basic tasks for image analysis,namely object detection and attributes analysis,which are improved signifi-cantly by deep learning methods.However,in the past,the visual object detection and attributes analysis are implemented in a multi-stage framework,which has some drawbacks:1)the inaccu-*rate object detection will cause the accumulative error of attributes analysis;2)it cannot capture the correlation between multiple tasks;3)the training and testing processes are complex.This paper proposes a multi-task deep learning based end-to-end object detection and attributes anal*ysis algorithm to improve the current multi-stage framework.The contributions of this paper:1.This paper proposes a novel deep learning based end-to-end object detection and at-tributes analysis algorithm based on Faster R-CNN[1]with the following advantages:1)it incor-porates the contextual information of the objects in the image and thus reduces the accumula-tive error caused by inaccurate detection;2)it achieves end-to-end joint learning by multi-task learning,including object detection and attributes analysis,which fully utilizes the correlation between labels of multiple tasks and boosts the generalization ability.The end-to-end model also simplifies the training and test processes,which improves the computational efficiency.2.This paper applies the end-to-end algorithm to gesture interaction tasks in egocentric view,including gesture detection,recognition and key points localization,all of which outper-form the multi-stage algorithm and validate the advantages of the end-to-end framework.3.This paper implements an Air-Writing-Recognition System in egocentric view based on the end-to-end gesture interaction algorithm.With the gesture category as the interaction command and the key points movement as the hand writing trajectory,the system recognizes the hand writing.It satisfies the requirement of good user interaction experience and validates the value of this algorithm.4.This paper applies the end-to-end algorithm to car license plate detection.It detects the car license plate accurately and obtains the directed car license plate by analyzing it's multi?direction attribute.It paves the way of subsequent car license plate recognition.