多对象图像数据集建立及显著性检测算法评估

图像视觉显著性检测算法在已有数据集上已经取得很好的结果,但是目前的多个数据集存在两个严重的问题:首先,数据集中的图像以只包含一个显著对象的图像为主;其次,在建立显著对象标注结果的过程中,忽略了用户对同一幅图像中包含的多个显著对象的不同认知。上述问题导致了在已有数据集上对显著性检测算法进行评估,不能体现算法在实际应用中的真实效果。为此,提出体现用户认知的多显著对象图像标注方法,首先设计并实现辅助软件,收集用户对各显著对象的重要程度的认知情况,包括显著区域与相应的重要程度;然后融合收集的多用户数据,绘制出以灰度图为表现形式的显著对象标注结果,并通过灰度值体现多用户对于每个显著对象的认知情况。基于改进的显著对象标注方法,建立了一个包含1000幅多显著对象图像的数据集,并为每幅图像提供了体现用户认知的显著对象标注结果。对10种具有代表性的显著性检测算法在已有数据集和建立的数据集上的性能进行了比较。实验结果表明,这些显著性检测算法在建立的数据集上的性能有大幅度的降低,例如受试者工作特征曲线下面积(ROC-AUC)评估参数的最大降幅超过了0.5,这证实了已有数据集存在的问题及建立新数据集的需求,同时指出显著性检测算法在处理包含多显著对象的复杂图像上存在的不足。

New multi-object image dataset construction and evaluation of visual saliency analysis algorithm

Image visual saliency analysis algorithms have achieved satisfactory performance on existing datasets, but these datasets have two major problems. Firstly, most of the images contain only one salient object. Secondly, users' cognition of multiple salient objects in the same image was ignored during building salient objects' ground truth. The above problems result in that the performance of saliency analysis algorithms used in the real applications cannot be reflected by the evaluation on the existing datasets. So in this paper, a novel method of labeling the ground truth of salient objects was proposed. Firstly, a software to collect users' cognition of the important values of multiple salient objects in each image was designed and implemented. Then, according to the collected data from each user, the ground truth map represented as a gray scale image was created by manually labeling the regions covered by the salient objects. The pixel value of each region equals to the collected saliency in the first step. Based on the improved ground truth labeling method, a salient object dataset contains 1000 multi-object images was built. A ground truth map for each image was created to record users' cognition of the objects' saliencies. Then 10 state-of-the-art saliency analysis algorithms on existing datasets and the established dataset were compared. The experimental results show that these algorithms' performances are greatly reduced on the established dataset, such as the Area Under Curve of Receiver-Operating Characteristic (ROC-AUC) has a greatest decline of more than 0.5. The results prove the problems of existing datasets and the demand of building a new dataset, and point out the insufficiency of saliency analysis algorithms on complex images with multiple salient objects.