Active Learning Sample Selection Method For Object Detection

As an important fundamental task in the field of computer vision,object detection has a wide range of applications and high value,from screening and diagnosis of diseases in medical images to detecting defects in industrial production and monitoring abnormal behaviors in security surveillance,all of which rely on the support of object detection algorithms.Through object detection,computers can automatically recognize,classify,and locate objects,thus achieving more intelligent applications.The performance of object detection not only depends on excellent network structures but also requires the support of large and rich annotated datasets.However,annotating data requires a considerable amount of personnel and funding,and the quality of the annotated data greatly affects the accuracy of object detection algorithms.To address the issue of annotation cost,researchers have proposed active learning methods for object detection.The purpose of this method is to select high-information samples from unlabeled datasets for manual annotation using a fixed sample selection strategy,providing higher information with fewer labeling costs,and helping object de-tection algorithms improve performance quickly.However,there are currently few active learning methods specifically for object detection.Most active learning algorithms are designed for image classification tasks and suffer from problems such as computational complexity and poor generality,which are not entirely applicable to object detection tasks.Therefore,this paper proposes the following two active learning algorithms specifically for object detection:Active learning method based on pseudo-labeling.A multi-scale feature fusion method is designed to fuse and reduce the dimensionality of the multi-scale features of unlabeled samples,while retaining the main features of the samples and reducing com-putational burden.Clustering is used to apply pseudo-label classification to unlabeled samples,and information entropy is introduced to calculate the uncertainty of unlabeled samples,considering both uncertainty and diversity of the unlabeled samples.Active learning method based on feature distance.A feature partitioning method is designed to partition sampling regions based on the feature distance between labeled and unlabeled samples.High information samples are introduced by sampling the uncertainty in the three sampling regions,while avoiding errors caused by the overconfidence of object detection algorithms,thus ensuring the diversity of the labeled dataset.