中文摘要：

候选框生成算法能够有效减少分类器所需处理的图像块个数,提高目标检测效率和准确率。二值化梯度范数（binarized normed gradients,BING）算法计算速度快,检测率高,但是生成的候选框平均召回率很低。在BING算法的基础上,引入基于图的超像素分割算法来提高平均召回率。实验结果表明,文章算法能够有效提升生成的候选框的平均召回率;综合多个参数值下的结果,该算法能够达到60%以上的平均召回率,而BING算法在使用大量候选框条件下的平均召回率仍低于36%。

英文摘要：

Proposal generation can reduce image patches to classify significantly and improve the object detection efficiency and precision. BING has the fast computation and high detection rate, but average recall of the proposals it generates is very low. Based on BING, the article induces graph- based segmentation （GS） algorithm to help improve the average recall. The experiment results show that the proposed algorithm can effectively improve average recall, and the combined results under several parameter values suggest the average recall of our algorithm can reach above 60% while the average recall of BING is still under 36% with large amount of proposals.