基于背景学习的显著物体检测

目的 显著物体检测的目标是提取给定图像中最能吸引人注意的物体或区域，在物体识别、图像显示、物体分割、目标检测等诸多计算机视觉领域中都有广泛应用。已有的基于局部或者全局对比度的显著物体检测方法在处理内容复杂的图像时，容易造成检测失败，其主要原因可以总结为对比度参考区域设置的不合理。为提高显著物体检测的完整性，提出背景驱动的显著物体检测算法，在显著值估计和优化中充分利用背景先验。方法 首先采用卷积神经网络学习图像的背景分布，然后从得到的背景图中分割出背景区域作为对比度计算参考区域来估计区域显著值。最后，为提高区域显著值的一致性，采用基于增强图模型的优化实现区域显著值的扩散，即在传统k-正则图局部连接的基础上，添加与虚拟节点之间的先验连接和背景区域节点之间的非局部连接，实现背景先验信息的嵌入。结果 在公开的ASD、SED、SOD和THUS-10000数据库上进行实验验证，并与9种流行的算法进行对比。本文算法在4个数据库上的平均准确率、查全率、F-measure和MAE指标分别为0.873 6、0.795 2、0.844 1和0.112 2，均优于当前流行的算法。结论 以背景区域作为对比度计算参考区域可以明显提高前景区域的显著值。卷积神经网络可以有效学习图像的背景分布并分割出背景区域。基于增强图模型的优化可以进一步实现显著值在前景和背景区域的扩散，提高区域显著值的一致性，并抑制背景区域的显著性响应。实验结果表明，本文算法能够准确、完整地检测图像的显著区域，适用于复杂图像的显著物体检测或物体分割应用。

Salient object detection based on background learning

Objective Salient object detection aims to identify spatial locations and scales of the most attention-grabbing objects in a given image, which is shown to be helpful in various computer vision tasks, such as object recognition, adaptive image display, and object detection. Different from eye-fixation saliency prediction, salient object detection emphasizes the saliency and wholeness of detected objects. Thus, dealing with cluttered background and diversity of object parts within an image has always been one of the major challenges in salient object detection. Bottom-up visual saliency is commonly characterized by the contrast of primitive image features at the pixel or super-pixel levels because contrast is the most predominant factor in human cognition. In the literature, the local or global contrast is usually adopted to straightforwardly derive the saliency map, where the contrast in a certain region is calculated by comparing its feature with that of the reference regions. However, such methods using local or global contrast reference regions may fail to detect whole salient objects, especially when dealing with complicated images. We attribute their failure to unreasonable setting of contrast reference regions. To enhance the wholeness of the detected salient objects, an explicit background-driven method is proposed, in which background prior is comprehensively utilized in saliency estimation and optimization. Method To obtain the background regions of an image for contrast estimation, deep convolutional neural networks were initially used to learn a background map representing the likelihood of each region belonging to the background. From the obtained background map, the background regions could be segmented with a simple thresholding strategy. The learned background regions were then used as references for region contrast computation. To enhance the consistency between the foreground and background regions, enhanced graph-based optimization was adopted to propagate saliencies along the graph. Besides the conventional local connections in a k-regular graph, prior connections with virtual nodes and non-local connections between nodes belonging to background regions were also added to the graph to embed the learned background prior. Result To verify the effectiveness of the proposed salient object detection method, comprehensive experiments were conducted on four public saliency detection datasets, namely, ASD, SED, SOD, and THUS-10000. The results were compared with those of nine state-of-the-art methods. Four indicators (i.e., precision, recall, F-measure, and MAE) were adopted for comparison. The average scores in precision, recall, F-measure, and MAE of our method were 0.873 6, 0.795 2, 0.844 1, and 0.112 2, respectively, which showed that our method outperformed other popular methods. The best results on all of the datasets were achieved using the proposed method, thereby demonstrating its effectiveness and superiority. Conclusion Restricting the contrast reference regions to the background could significantly improve contrast-based saliency estimation. The background regions in an image could be effectively learned by convolutional neural networks. An enhanced graph-based optimization could fuse the saliency confidences of different parts from the same object to discover the whole salient object; thus, a more consistent and background-suppressed saliency map could be generated. Experimental results showed that the proposed method can be successfully used in salient object detection and object segmentation in natural images.