基于结构感知深度神经网络的显著性对象检测算法

Salient object detection algorithm based on structure-sensitive deep neural network

由于现有的基于深度神经网络的显著性对象检测算法忽视了对象的结构信息,使得显著性图不能完整地覆盖整个对象区域,导致检测的准确率下降。针对此问题,提出一种结构感知的深度显著性对象检测算法。该算法基于一种多流结构的深度神经网络,包括特征提取网络、对象骨架检测子网络、显著性对象检测子网络和跨任务连接部件四个部分。首先,在显著性对象子网络的训练和测试阶段,通过对象骨骼检测子网络学习对象的结构信息,并利用跨任务连接部件使得显著性对象检测子网络能自动编码对象骨骼子网络学习的信息,从而感知对象的整体结构,克服对象区域检测不完整问题;其次,为了进一步提高所提方法的准确率,利用全连接条件随机场对检测结果进行进一步的优化。在三个公共数据集上的实验结果表明,该算法在检测的准确率和运行效率上均优于现有存在的基于深度学习的算法,这也说明了在深度神经网络中考虑对象结构信息的捕获是有意义的,有助于提高模型准确率。

Current salient object detection algorithms based on deep neural network (DNN) are usually not able to be aware of the structure of instance, making the generated saliency maps fail to cover the entire salient object region,thus drag down the accuracy. To solve this problem, this paper introduced a novel multi-stream deep neural network, which integrated four components in a single framework, feature extractor, object skeleton sub-network, salient object sub-network and cross-domain connections. Firstly, during the learning and testing process, the salient object detection sub-network encoded the object structure which was extracted by using object skeleton detection sub-network through the cross-domain connections, so as to make the deep model be aware of the information of object structure and overcome the problem of incomplete detection of the target area. Then, to further improve the accuracy, it used a dense conditional random field-based algorithm as the refinement post-process, so as to generate a more accurate saliency map as the results. Experimental evaluations were conducted on three widely-used benchmarks.The results show that the proposed algorithm outperforms all existing DNN-based detection algorithms in accuracy and efficiency. It also indicates that integrating object structure information into deep neural network model is meaningful, which can help to improve the overall accuracy.