CAJ | 학술논문

为了实现重叠苹果目标的精确分割，提出了一种 Snake 模型与角点检测相结合的重叠苹果目标分割方法。该方法首先利用 Snake 模型得到重叠苹果目标的轮廓，接着采用提出的基于距离测度的角点检测算法寻找重叠苹果目标的角点，针对距离扰动产生伪角点的问题，采用3级db1小波变换得到不含细节信号的近似距离信号，并通过Spline样条内插算法使其恢复到原始信号的长度，从而去除伪角点，最后提出了一种基于长轴分割准则的分割点选取方法并实现了重叠苹果目标的分割。为了验证算法的有效性，利用20幅重叠苹果目标进行了试验，并与人工计算得到的分割线进行了对比，试验结果表明，利用文中算法分割重叠苹果目标的最大误差为13.27°，最小误差为1.20°，平均误差为6.41°，表明Snake模型与角点检测算法相结合对重叠苹果目标具有较好的分割性能，将该方法应用于重叠苹果目标的分割是可行的。
위료실현중첩평과목표적정학분할，제출료일충 Snake 모형여각점검측상결합적중첩평과목표분할방법。해방법수선이용 Snake 모형득도중첩평과목표적륜곽，접착채용제출적기우거리측도적각점검측산법심조중첩평과목표적각점，침대거리우동산생위각점적문제，채용3급db1소파변환득도불함세절신호적근사거리신호，병통과Spline양조내삽산법사기회복도원시신호적장도，종이거제위각점，최후제출료일충기우장축분할준칙적분할점선취방법병실현료중첩평과목표적분할。위료험증산법적유효성，이용20폭중첩평과목표진행료시험，병여인공계산득도적분할선진행료대비，시험결과표명，이용문중산법분할중첩평과목표적최대오차위13.27°，최소오차위1.20°，평균오차위6.41°，표명Snake모형여각점검측산법상결합대중첩평과목표구유교호적분할성능，장해방법응용우중첩평과목표적분할시가행적。
To achieve successful segmentation of overlapped apples, a segmentation method by using Snake model and corner detectors was presented. As contour is an important basis for detection and recognition of object, and remarkable characteristic of overlapped apples has some typical angular points, which are also called segmentation points and in the target contour. Since Snake model could better converge to target’s concave places, Snake model was used to extract overlapped apples’ outline. For searching overlapped apples’ corner points, corner detection algorithm based distance was proposed:1) overlapped apples’ contour was coded;2) the distance between contour points and the given‘center point’ was calculated, where‘center point’ was overlapped apples’ centroid point for the simplicity of calculation;3) the distance curve that was get in step 2 is useless as it may engender a lot of spurious corner points. This is caused by small disturbances of small distance, for removing spurious corner points, db1 wavelet was utilized to decomposed original signal at level three, there is a relationship between wavelet transform and digital filter banks. so the wavelet transform can be simply achieved by a tree of digital filter banks. The idea behind filter banks is to divide a signal into two parts:one is the low frequency part and the other is the high frequency part, which could be achieved by a set of filters, the low frequency that is approximate version of the original distance curve in this paper don’t contain detail components of original distance and is beneficial to detect true corner points. But the problem with the use of these filters is that each of the two decomposed signals is subjected to downsampling, which simply means throwing away every second data point. After decomposition with three levels, the length of approximated signal reduced, which may cause the miss of the index of original contour point. As for this reason, the approximated signal must be recovered to its original length. In this paper, the interpolation algorithm with the use of splines is carried out to recover the length of approximated signal;4) corner points represent the enormous changes of curvature and is shown the maximum or minimum on the distance curve, therefore, corner points can be detected by calculating the extrimum of distance curve. Detected corner points need to be selected to determine overlapping positions, a segmentation method based long axis segmentation rule was proposed to choose segmentation line:1) overlapped apples were divided into uniform two parts approximately by long axis;2) segmentation line was chose by calculating the distance between bilateral corner points and centroid point. Some split criteria were given as:1) the direction of the detected points should be opposite, which meant that the detected points from the same region should not be used to split an object; 2) the length of split line should be short. By using these given criteria, the detected corner points were matched to realize the segmentation of overlapped apples. To validate the effectiveness of the algorithm, 20 overlapped apples in nature scenes were tested. Compared with segmentation line obtained by artificial calculation, highest segmentation error of the proposed method is 13.27°, minimum error is 1.20°, and the average error 6.41°. Experimental results show that the proposed segmentation algorithm has a preferable performance, and it is feasible and valid for overlapped apple segmentation in nature scenes.