一种针对特定目标的提议算法

Improved Edge Boxes Proposal Used in Specific Instance

Edge boxes是一种一般目标提议算法,为了提高其对特定目标提议的有效性,提出了一种改进的Edge boxes提议算法。Edge boxes算法通过对候选框评分得到候选区域,针对特定目标,改进了评分的依据。首先,提取目标的不同区域,得到被候选框完全封闭包围的轮廓在这些区域中的像素点,由各个区域的像素点的均值幅值构成特征向量;其次,计算各个候选框与目标的特征向量的差值,并将差值与被候选框完全封闭包围的轮廓的数量结合起来,作为评价各个候选框包含目标可能性大小的依据。所提算法在视频序列上进行了实验,精度至少提升了10.71%,验证了改进算法的有效性。

Since edge boxes cannot accurate in detecting specific instance,an improved detection algorithm is proposed.Edge boxes evaluate boxes by calculate the amount of the completely closed contours in the boxes,as for specific instance,we make some improvement in the process of scoring.Firstly,we extracted the different regions of the target,obtained the completely closed contours in the boxes and the pixels that compose these contours.Based on what we extracted and what we obtained,a feature vector composed by the mean magnitude of these pixels we obtained in the different regions is proposed.Then,we need to calculate the differences between a candidate and the target in the feature vector.Finally,combining the differences and the amount of the completely closed contours in the boxes,the probability of the target in the boxes is known.The proposed algorithm is tested on the video sequence.Experimental results show that has a significantly improvement of 10.71%in representative recall rate.The improved algorithm is effective for improving detecting performance.