Semi-automatic Annotation Method of Infrared Sequence Images Based on Image Significance
Chuang Liu, Pan Huang, Xiaogang Yang, Ruitao Lu · 2021 IEEE 3rd International Conference on Frontiers Technology of Information and Computer (ICFTIC) · 2021
At present, there are some problems in manually labeling data, such as slow speed, low efficiency and low precision. In order to improve the labeling efficiency, according to the characteristics of infrared sequential images, a semi-automatic labeling method of infrared sequential images based on image saliency is proposed. The kernel correlation filtering algorithm is used to track the infrared time series target, obtain the approximate area of the target, and realize the target position pre labeling. The saliency target detection algorithm is used to calculate the target saliency map of the pre-labeled area, and the foreground target is extracted by threshold segmentation, so as to realize the target fine labeling of the pre-labeled area. The average overlap rate with the manual annotation results is about 95%, which can accurately locate the target position and improve the annotation efficiency of sequence images.