Data annotation correction method based on frame regression

Abstract

A data annotation correction method based on frame regression comprises the following steps: dividing data into two batches of sample data, which are gold annotations and hard annotations, according to the difficulty degrees in the first annotation process and the confidence coefficients of an annotation result; using a focus loss function to improve a target detection algorithm YOLO V5, then using gold annotation sample data for training, and after training of the training models is stable, saving m training models every fixed iteration times; reasoning the stored m training models on the hard annotation sample data, and performing offline storage on all pictures formed according to the reasoning result; for each picture, summarizing all reasoning results of the m training models, clustering all frames, and the number of clusters being set to be the number of real targets on the current picture; counting the number of frames, and carrying out general distribution modeling on four boundary points of all frames in the same cluster; and correcting the position of the frames according to a modeling result.

Description

technical field [0001] The invention relates to the technical field of deep learning, in particular to a method for correcting data labeling based on frame regression. Background technique [0002] At present, artificial intelligence technology with deep learning as the core has made breakthroughs in the fields of industrial vision, natural language processing, and automatic driving. Among them, in the field of industrial quality inspection, the precision of convolutional neural network for defect classification has surpassed that of human eyes, and the recognition speed of defects is far faster than that of human beings. The detection scheme and equipment have entered the stage of industrialization. [0003] Deep learning is an algorithmic weapon in the era of big data. It has algorithmic performance that is difficult to surpass by traditional machine learning. However, deep learning relies heavily on training data. In actual industrial scenarios, it is difficult to obtai...

AI Technical Summary

Problems solved by technology

[0005] According to the above analysis, the following problems still exist in the field of deep learning data labeling: 1. The workload of data labeling and correction is huge and the efficiency is low; 2. The difficulty of labeling difficult samples is high, and the difference in the subjective consciousness of labelers will lead to inconsistencies in labeling 3. The low-quality labeling of difficult samples has a negative impact on algorithm training

Images