The Research And Implementation Of Object Detection Model Based On Deep Learning And Crowdcomputing

The purpose of object detection based on deep learning is to use the technology and method of deep learning and image processing to locate all objects of interest in the image and determine their semantic categories.Deep learning can greatly improve the accuracy of object detection,which mainly depends on more and more neural network layers and large training datasets.However,large datasets need to spend a lot of manpower and material resources to collect.The lack of large labeled datasets has become the development bottleneck of deep network models in many applications.Crowdcomputing,which collects a large amount of annotation information by gathering crowd intelligence,has a natural advantage in solving the problem of datasets construction.However,It often cannot guarantee the quality of annotation.It is a better choice to combine with ground truth inference algorithm.The existing ground truth inference work is not suitable for the current popular object detector based on the anchor mechanism.Object detection is a kind of task combining classification and regression.At present,there is no deep learning object detector trained with repeated crowdsourcing labels training.Considering the challenges and problems above,this paper concentrates on the research and analysis of the combination of object detection technology and ground truth inference algorithms.We design and implement an efficient crowdsourcing annotation process to complete the deployment of the whole pipeline of the object detection system.The main research contents are as follows:1)designing multi-objective image annotation process based on crowdsourcing and researching a reasonable process to complete the annotation task for the subsequent learning algorithm.In order to continuously and efficiently collect the repeated crowdsourcing labels,this paper proposes a scheme to divide the annotation task into multiple subtasks in the multi round annotation mechanism,and solves the problems of difficult annotation,uncontrollable annotation quality and unique label output in the crowdsourcing way with the assistance of training and verification links;2)building the object detection model trained with repeated crowdsourcing data.designing the ground truth inference algorithm of object detection labels and combining it with the two-stage object detector to complete the end-to-end training.Using repeated labels to reduce the regression error of the bounding box and the classification error of the annotated object category,improve the label quality,and improve the accuracy of the object detection system based on crowdsourcing labels end-to-end;3)Designing and implementing the object detection system based on crowdcomputing and deep learning.Based on the above two research contents,the object detection system is built,and the core functions such as data annotation,model training and inference API call are realized.This paper develops and implements a deep learning object detection system based on the above research achievements,which can greatly reduce the time consumption of data collection and accelerate the application of object detection.On the other hand,the object detection system can make users use of high-precision object detection technology to meet business requirements.It can be applied to general scenarios or application scenarios in specific fields and improve user experience through feedback learning.