针对Edge Boxes算法召回率不高的问题, 并结合目标的显著性检测, 提出了一种基于颜色距离与Edge Boxes候选区域算法。首先利用结构化边缘检测算子获取图像的边缘特征, 并通过边缘点聚合及边缘段相似性策略, 获取每个边缘段的权值; 其次, 在待检测图像上无重叠采样若干图像块, 记作C图像块, 并将C图像块向周边延拓像素, 获取S图像块; 然后, 根据颜色直方图, 计算两图像块各颜色通道的卡方距离, 并赋予合适权重作为该C图像块的显著性得分; 最后, 统计滑动窗口内边缘段的数量和C图像块数, 确定候选区域。在PASCAL VOC 2007验证集上实验, 当交并比取05, 0.6, 0.7, 候选区域个数为2 000时, 与Edge Boxes相比, 所提算法的召回率分别提高了0.46%, 0.35%, 0.57%。每张图像的运行时间大约为0.43 s, 这表明, 所提算法以牺牲微小计算资源却能够有效改善候选区域质量。

For the low recall rate of Edge Boxes algorithm, a region proposals algorithm based on color distance and Edge Boxes is proposed in this paper, combining with the object saliency detection. Firstly, the edge features are computed using Structured Edge detector, and the weight values of every edge group are obtained through edge point aggregation and edge groups affinities. Secondly, several image patches are sampled randomly in the image, which is to be detected, this image patches are denoted C image patches and the S image patches are obtained through expanding the pixels of C image patches. The Chi-square distance of this two image patches are computed according to the histogram of each channel of C and S image patches, the salient scores of this image patch are obtained with putting appropriate weight value on chi-square distance of three channels. Finally, the region proposals are determined with the number of edge groups and C patches in sliding windows. The experimental results in PASCAL VOC 2007 validation set show that the object recall of the proposed algorithm increases 046%, 0.35%, 0.57% than Edge Boxes at intersection over union threshold of 0.5, 0.6, 0.7 in 2 000 region proposals, respectively. Our approach runs about 0.43 s in an image, this demonstrates that the proposed algorithm can effectively improve the quality of region proposals with only minor loss in computation resources.