快速优化筛选多尺度矩形域的二进制描述

发布时间：2018-04-19 02:14

本文选题：目标识别 + 特征描述　；参考：《中国图象图形学报》2016年03期

【摘要】：目的为更好地兼顾基于手动设置的二进制特征描述子优越的实时性能和基于优化学习的二进制特征描述子鲁棒的区分性能,提出一种快速优化筛选多尺度矩形域的二进制描述算法(MRFO),运用于识别卫星装配时所需的典型工件目标。方法按像素的灰度值和梯度方向划分图像并利用不同的高斯核函数进行平滑,建立多尺度的子图像集合;从多尺度的子图像中,快速通过约束条件提取候选矩形域;在训练阶段,通过优化学习计算候选矩形域的相关得分及最优阈值,筛选出其中具有强区分性和低相关性的集合;在测试阶段,计算筛选出的矩形域响应值并利用最优阈值进行二值化,将结果依次串联构成二进制描述向量。结果实验通过ROC曲线图和80%精确率条件下的召回率统计结果证明MRFO描述算法具有优越的区分性能,平均的精确度能够高出对比算法8%~12%;并在真实的视频图像中利用MRFO描述算法识别出典型工件目标;根据训练阶段的执行时间只有传统优化学习算法的4.35%,只是在测试阶段略高于手动设置的二进制描述算法,证明MRFO描述算法具有优良的实时性能。结论 MRFO描述算法能够更好地克服各种视角、尺度和旋转变换的干扰以及周围相似背景信息的影响,准确识别出典型工件目标,有助于提高卫星的地面装配精度和效率,改善国内相关行业的自动化水平。普遍适用性较强,具有良好的应用前景。
[Abstract]:For the purpose of better real-time performance based on both binary feature descriptor set manually and superior performance robust feature descriptor based on differentiated binary optimization study, we propose a fast binary optimization screening multi-scale rectangular domain description algorithm (MRFO), applied to the typical workpiece required during assembly. The satellite recognition method according to the pixel the gray value and gradient direction image partition and smoothing Gauss using different kernel functions, set sub image using multi scale; from the multi-scale sub image, fast constraints through the extraction of candidate rectangle; during the training phase, learning scores and the optimal threshold by computing the candidate rectangular domain optimization, which are selected from has a strong distinction and set low correlation; in the test phase, calculation of rectangular domain selected response value and optimal threshold for binarization, the results Which are connected in series to form a binary vector description. Experimental results by ROC curve and 80% accuracy under the condition of the recall rate statistical results show that MRFO algorithm has superior performance to describe the distinction, the average accuracy of high contrast 8%~12% algorithm; and in real video images in using MRFO description algorithm to identify typical workpiece according to the training objectives; the execution time is only 4.35% of the traditional optimization algorithm, only in the binary test phase is slightly higher than the manual setting description algorithm, prove that MRFO algorithm has good performance in real-time is described. Conclusion the MRFO description of the algorithm is better able to overcome the influence of various angle, rotation and scale transform and interference around similar background information, accurately identify typical parts, help to improve the accuracy and efficiency of satellite ground assembly, improve the automation level of the domestic industry generally. It has strong applicability and good application prospect.

【作者单位】：河北工业大学机械工程学院;
【基金】：国家高技术研究发展计划(863)基金项目(2015AA043101)~~
【分类号】：TP391.41

【相似文献】