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。

Objective To better balance binary feature descriptor algorithms based on manual learning, which have superior real-time performance, and binary feature descriptor algorithms based on optimization study, which have robust performance, this paper presents a binary feature description algorithm for fast optimization screening of a multi-scale rectangular area (referred to as MRFO). The typical workpiece target in satellite assembly is identified. Method The proposed description algorithm divides images according to the pixel gray value and gradient direction and simultaneously smoothens each sub-image with different Gaussian kernel functions to establish a multi-scale image set. Candidate rectangular areas are rapidly extracted through the constraint condition from the sub-image of multi-scale or the multi-scale feature point neighborhood. In the training phase, the proposed algorithm calculates the score and optimal threshold of candidate rectangular areas by using optimization study and selects the subset that has strong distinction and low correlation. In the testing phase, the proposed description algorithm calculates the response value of selected rectangular areas in the multi-scale feature point neighborhood and employs the optimal threshold for binarization to constitute the binary description vector of feature points. Result The experiment proved that the binary feature description algorithm for fast optimization screening of a multi-scale rectangular area demonstrates a robust performance based on the ROC curve and the recall rate statistical result under the condition of 80% accuracy rate. The average accuracy is higher by 8% to 12% than that of compared algorithms. In real video images, the proposed description algorithm can identify the typical workpiece target accurately. The experiment also proved the superior real-time performance of the proposed algorithm; the execution time in the training phase is only 4.35% of the traditional optimization learning algorithm (only slightly higher than that of the binary description algorithm based on manual learning in the testing phase). Conclusion Compared with the traditional scheme of binary feature description algorithms and float feature description algorithms, the proposed feature description algorithm can overcome interference from all types of perspective, scale, rotation transformation, and influence of similar background information. The proposed algorithm can identify the typical workpiece target accurately. The algorithm can help improve the accuracy and efficiency of satellite assembly and enhance the automation level of domestic related industries. The universality of the method is strong, and it has a good application potential.