基于Faster R-CNN的显著性目标检测方法

显著性目标检测成为计算机视觉领域中的研究热点问题之一,但目前的方法在面对前景和背景对比度不强及复杂背景的图像时,较难取得好的检测效果。融合多尺度超像素分割方法,提出一种在背景信息相对复杂的场景中基于FasterR-CNN的显著性目标检测方法。首先对图像进行多尺度超像素分割,同时利用FasterR-CNN对图像进行目标检测,根据似物性特点对超像素进行显著性筛选,得到初始目标位置特征后进行显著性检测及优化,最后使用元胞自动机方法对多尺度超像素显著性图进行融合。通过在特定类数据集进行实验,与已有典型显著性检测进行对比分析,验证了本文方法在背景复杂的图像中可提升显著性目标检测的精度。

Salient Object Detection Based on Faster R-CNN

Saliency detection becomes an important topic in computer vision. However,most of existing approaches often fail when the contrast between foreground and background is similar or in complex background. We propose a salient object detection based on Faster R-CNN. It extracts the salient object in the background with relatively complex information by combining multi-scale superpixel segmentation. Firstly,we segment the input image using multi-scale superpixel segmentation and detect the object by Faster R-CNN. Secondly,according to the objectness of the object,the salient superpixels are selected and obtain the original position of the object,and the salient map is highlighted and optimized. Lastly,we use Multi-layer Cellular Automata(MCA) to fuse the multi-scale superpixel salient maps and gain the final salient maps. We evaluate the proposed approach on specific dataset and compare with the state-of-art method,which prove our approach can improve the salient object detection accuracy with the complex background and clutter scene.