摘 要: 随着近年来深度学习技术的发展,图像显著性目标检测的研究重点偏向于利用深度学习方法解决问题。为了全面且深入地探究图像显著性目标检测领域,基于深度学习框架,回顾近五年出现的20余种深度学习方法,归纳出四类深度学习策略,并且对比了它们在4 个显著性数据集上的评价结果。结果显示,各类策略在不同数据集上的F度量值为0.800—0.950,综合利用多种策略的方法可以取得更优的预测指标,但仍然存在复杂场景干扰下检测有误的问题。针对现有问题,提出加强深度学习方法在复杂数据集上的训练,进而优化显著目标预测结果的定位准确性及边缘完整性。 |
关键词: 图像显著性目标检测;深度学习框架;深度学习策略 |
中图分类号: TP391
文献标识码: A
|
|
Overview of Image Salient Object Detection Research based on Deep Learning |
LI Yuanzhen1, ZHAO Junsong2
|
( 1.School of Artificial Intelligence, Hebei University of Technology, Tianjin 300401, China ; 2.Information Network Center, Civil Aviation University of China, Tianjin 300300, China)
2081059521@qq.com; xmbdhyq@163.com
|
Abstract: With the development of deep learning in recent years, research focus of image salient object detection tends to use deep learning methods to solve problems. In order to comprehensively and deeply explore the field of image salient object detection, this paper proposes to review more than 20 deep learning methods over the past five years based on deep learning framework, summarize four types of deep learning strategies, and compare their evaluation results on four datasets of salient object detection. The results show that F-measure of various strategies on different datasets are from 0.800 to 0.950. The comprehensive use of multiple strategies can achieve better prediction indicators, but there are still the problems of incorrect detection under the interference of complex scenes. In view of the existing problems, it is proposed to strengthen the training of deep learning methods on complex datasets, so as to optimize the positioning accuracy and edge integrity of salient object. |
Keywords: image salient object detection; deep learning framework; deep learning strategies |