Sample Acquisition Automatic Labeling And Data Enhancement System Applied To Deep Learning

In recent years,target detection algorithms based on convolutional neural networks in the field of computer vision have developed rapidly.Compared with traditional target detection methods,the emergence of some classic deep learning target detection networks has greatly improved the accuracy and speed of target detection,laying a solid foundation for target detection in practical applications.However,because of the characteristics of the large number of parameters and the deep layers in convolutional neural networks,in order to make the network fully learn,train,and obtain better detection results,it is necessary to construct a large-scale training data set for it.But in practical applications,due to the specificity and uniqueness of each object to be detected in the detection scene,it is necessary to construct large-scale data sets through laborious manual annotation for the objects to be detected.Although there are some classic data enhancement methods that can expand the scale of manually labeled data sets,the ordinary manual shooting and manual labeling process limits the enhancement methods in the target detection data set to only a single sample enhancement,such as rotation,blur,or other methods.Aiming at the various problems in the above practical applications,this paper proposes and implements a target detection network application system with small original data sample that improves the deployment efficiency and detection quality of the network in practical applications.In view of the lack of data sets,manual labeling time and high labor costs,the workflow of the classic deep learning network in practical applications has been optimized from three links:data collection,data labeling,and data enhancement,reduced the time consumption required for each step in network application.And by automating the steps,the degree of manual participation in network applications is reduced,and also the deployment efficiency of the network system is improved.The system first starts from the 3D scanning modeling method,performs 3D modeling of the object to be detected and generates a full-scale feature image of the corresponding object.After that,through the aliasing data enhancement step based on the contour of the object,it has been effectively enhanced to increase the size and information of the data set,so that the network can fully learn the target to be detected and achieve a certain detection result.Afterwards,in view of the defects of the 3D scanning modeling method,such as low texture resolution and the inability to effectively collect data on small objects,the method of surround video shooting was also proposed for data collection and corresponding enhancement experiments,and comparative analysis The characteristics of the two collection methods are described.After accelerating and optimizing different steps in the system,this article analyzes the work efficiency and detection performance of the system from various angles at the end,verified that this system can accelerate network deployment and improve the effect of object detection in practical applications,and then promote the landing process of deep learning networks in practical applications.