Research On Active Semi-supervised Learning Target Detection Algorithm

With the increasing maturity of deep learning technology,the target detection algorithm based on deep learning has made remarkable achievements.However,the improvement of algorithm performance depends heavily on large-scale label samples.Although it is currently in the era of big data,labeled samples are still scarce,and a large number of unlabeled samples are rarely used.How to improve the utilization of existing labeled samples and how to mine unlabeled sample information to directly participate in model training is the subject of this article.Active learning calculates the amount of unlabeled sample information,screens out the unlabeled samples that are the most effective for the convergence of the current model's decision boundary,performs expert labeling and participates in model training,thereby improving labeling efficiency.Semi-supervised learning directly mines unlabeled sample information,and calculates semi-supervised learning loss directly participates in model training.Active learning and semi-supervised learning have complementary advantages and complement each other.This thesis focuses on the application of active semi-supervised learning in target detection.First,this thesis analyzes the models of active learning and semi-supervised learning respectively,and proposes an active semi-supervised learning framework based on unlabeled sample pools.The validity of the framework is verified by using image classification which is one level lower than the difficulty of target detection.Combining the uncertainty-based active learning image classification algorithm and the semi-supervised learning image classification algorithm based on consistency regularity,an active semi-supervised learning image classification algorithm based on stability is proposed.The experiment fully proves the effectiveness of the active semisupervised learning framework proposed in this thesis,and uses it as the basic framework for subsequent research on active semi-supervised learning target detection.Secondly,because the target detection sample label needs to be described in the two dimensions of classification and positioning,the active semi-supervised learning strategy for target detection needs to be studied in the two dimensions of positioning and classification.An active learning target detection algorithm based on uncertainty is proposed.In this algorithm,the evaluation indicators of unlabeled sample information include classification uncertainty,positioning tightness,and positioning stability.After that,a semi-supervised learning target detection algorithm based on consistency regularity is proposed.The algorithm introduces consistency classification loss and consistency location loss to mine unlabeled sample information directly to participate in model training.Finally,this thesis combines the active learning target detection algorithm based on uncertainty and the semi-supervised learning target detection algorithm based on consistency regularity.In the active semi-supervised learning framework based on the unlabeled sample pool,an active semi-supervised learning framework based on twodimensional stability is proposed.Supervised learning target detection algorithm.Experiments prove that the active semi-supervised learning target detection algorithm based on two-dimensional stability proposed in this thesis can effectively save the number of label samples by 10%?32% under the premise of obtaining the same performance.