Research On Interactive Intelligent Assistant Annotation Algorithm

In recent years,the rapid iterative progress of supervised deep learning technology has greatly promoted the development of artificial intelligence applications^[1].The development of supervised deep learning relies heavily on large-scale labeled data sets^[2],but because manual labeling of data sets is expensive and time-consuming^[3],data set labeling errors are inevitable,and domain-oriented expertise is required when labeling.The quality of the original data is uneven,and these problems seriously restrict the production of high-quality labeled data.Therefore,the labeling of data sets for specific application fields has strong practical significance and research value,and is an urgent problem to be solved.In response to the above problems,this paper develops a labeling system for specific application domain data sets.First,in order to improve the efficiency and accuracy of manual marking,an image annotation system is proposed.This system uses a recommendation algorithm to recommend the most similar pictures to the marking personnel based on the pictures selected by the marking personnel,and iteratively according to the marking of the marking personnel It is recommended until all the labeling work is completed.A background framework that uses the labeling algorithm proposed in the subsequent chapters to recommend pictures,assists manual labeling,and finally implements image-level labeling of pictures.Secondly,with the rapid development of the image synthesis technology based on the adversarial generation network,the generated data is more and more real,so the authenticity of the collected data needs to be tested during data cleaning^[4].Taking the cleaning of face data as an example,this paper proposes a synthetic image detection network that integrates face detection and attention mechanisms for the diversity of face positions and sizes in the image.Without changing the input of the full image,the face detection network obtains a saliency feature map(face map)containing the face area based on the image to be detected,and merges the feature map output by the backbone network to enhance the human The features of the face area suppress the features of the non-human face area and improve the detection capability of the synthesized image.The experimental results show that the proposed method can accurately identify the synthesized samples and ensure the authenticity of the data set to be annotated.Secondly,in interactive annotation,for the problem of low recommendation efficiency of random and cosine distance similarity,the face features extracted by Face Net^[5]network are proposed,and the structure of"visual confusion map"is established based on LMNN metrics.The"picture"system recommends the most visually confusing images to the label for classification and labeling to improve the efficiency and accuracy of face classification and labeling.Finally,when performing image pixel-level labeling,using the interactive labeling method based on Deep Lab-V2^[6]architecture extreme point,the annotator enters the four extreme points of the object(top,bottom,left and right pixels)as The annotated guidance information adds an additional channel to the image in the input^[7]of the convolutional neural network(CNN),which contains the Gaussian function value centered at each extreme point,obtained with less user input Better annotation effects than classic interactive pixel-level annotation methods(such as Grab Cut,Random Walker,and i FCN).