Research And Implementation Of Object Detection Service Platform Based On Deep Learning

With the rapid popularization of intelligent terminals and the vigorous development of social media,the digital images shared on the Internet have exploded.Images bring great convenience to people's life and provide information as well as great challenges.The processing of digital images has become an important research subject in the fields of computer science,engineering,biological science,medicine and even social science.Object detection not only plays an important role in object recognition in real life scenes,but also supports many upper-level applications,such as intelligent video surveillance,robot navigation,image understanding and augmented reality,which have broad application prospects.The current problems of the application of object detection are:(1)There is still much room for improvement in the accuracy of the object detection algorithm,especially for small object detection;(2)Although the object detection algorithm can meet the real-time requirements in terms of speed,how to process the massive data more quickly still needs to be improved;(3)At present,most of the image service platforms at home and abroad are related to image classification and face recognition,but lack the service platform related to object detection.In view of the above problems,this paper focuses on the research and analysis of object detection technology based on deep learning,and realizes the design and development of object detection service platform.The main research contents include:(1)The object detection algorithm based on region proposal is optimized and improved with a new algorithm,called Faster R-CNN++.Based on the Faster R-CNN algorithm,the combination of hard example mining and multi-scale feature extraction mechanism are proposed.The mean average precision of the algorithm is improved by 8.5%on the PASCAL VOC 2007 dataset,the effect of small object detection has also been greatly improved;(2)The object detection algorithm based on regression is optimized and improved with a new algorithm,called SSD++.Depthwise separable convolution and model cropping technique are adopted on the basis of SSD algorithm,which increases the detection speed by 2.7 times.Detecting a single picture only takes 5.7ms,which not only can support real-time detection but also can quickly process massive images detection;(3)Propose the application scheme of object detection API service platform to solve the problem of lacking open service platform related to object detection at home and abroad.Based on the above research contents and achievements,this paper builds and implements an object detection service platform based on deep learning.This platform enables developers to accomplish the corresponding business requirements based on the object detection API designed in this paper.Both general objects and specific objects of the actual scene can be well applied.On the other hand,the object detection service platform can effectively solve the problem of missing datasets,and feedback learning can improve the performance of detection algorithm and enhance the user experience and the practical value of product.