基于子网络级联式混合信息流的显著性检测

Saliency detection hybrid information flows based on sub-network cascading

针对现有显著性检测算法在复杂场景下细节特征丢失的问题,本文提出了一种多层子网络级联式混合信息流的融合方法。首先使用FCNs骨干网络学习多尺度特征。然后通过多层子网络分层挖掘构建级联式网络框架,充分利用各层次特征的上下文信息,将检测与分割任务联合处理,采用混合信息流方式集成多尺度特性,逐步学习更具有辨别能力的特征信息。最后,嵌入注意力机制将显著性特征作为掩码有效地补偿深层语义信息,进一步区分前景和杂乱的背景。在6个公开数据集上与现有的9种算法进行对比分析,经实验验证,本文算法运行速度可达20.76帧/秒,并且实验结果在5个评价指标上普遍达到最优,即使对于挑战性很强的全新数据集SOC。本文方法明显优于经典的算法,其测试结果F-measure提升了1.96%,加权F-measure提升了3.53%,S-measure提升了0.94%,E-measure提升了0.26%。实验结果表明,提出的模型有效提高了显著性检测的正确率,能够适用于各种复杂的环境。

In view of the detail feature loss issue existing in the complex scenario of existing saliency detection algorithms,a fusion method of multi-layer sub-network cascade hybrid information flows is proposed in this paper.We first use the FCNs backbone network to obtain multi-scale features.Through the multi-layer sub-network layering mining to build a cascading network framework,the context information of the characteristic of each level is fully used.The detection and segmentation tasks are processed jointly.Multi-scale features are integrated by hybrid information flows,and more characteristic information with discernment is learned step by step.Finally,the embedded attention mechanism effectively compensates the deep semantic information as a mask,and further distinguishes the foreground and the messy background. Compared with the existing 9 algorithms on the basis of the 6 publicdatasets, the running speed of the proposed algorithm can reach 20.76 frames and the experimental results aregenerally optimal on 5 evaluation indicators, even for the challenging new dataset SOC. The proposed method isobviously better than the classic algorithm. Experimental results were improved by 1.96%, 3.53%, 0.94%, and 0.26%for F-measure, weighted F-measure, S-measure, and E-measure, respectively. These experimental results showthat the demonstrating the proposed model has higher accuracy and robustness and can be suitable for more complexenvironments, the proposed framework improves the performance significantly for state-of-the-art models on aseries of datasets.