基于Cascade Adaboost分类器的马铃薯快速定位方法

摘要

Aiming at the problem of difficultly locating potato with uneven surface grayscale and complex texture, and of easily misjudging the background area as potato area, this study put forward improvement based on a Cascade Adaptive Boosting (Cascade Adaboost) classifier potato positioning method and a candidate regional secondary screening method. The results showed that, after the optimization, the trained Haar+Cascade Adaboost classifier, LBP+Cascade Adaboost classifier and HOG+Cascade Adaboost classifier were used to test sets of potato image, the TP rate (true positive rate, detection rate), the FP rate (false positive rate, false alarm rate) and the overall recognition rate tested were 1.7％-0.8％-97.2％, 95.9％-0.0％-98.9％, and 86.7％-3.5％-93.9％, and the elapsed time was 8.2 ms, 7.5ms and 30.3ms, respectively. The Cascade Adaboost classifier could be used to quickly and precisely locate the target potato in movement, in which the LBP+Cascade Adaboost classifier is the best.%针对马铃薯表面灰度不均匀、纹理复杂不易定位的问题,通过采集类Haar(Haar-like)、局部二值模式(Local Binary Pattern,LBP)和方向梯度直方图(Histogram of Oriented Gradient,HOG),提出了基于级联自适应提升(Cascade Adaptive Boosting,Cascade Adaboost)分类器的马铃薯定位方法.同时,针对背景区域易误判为马铃薯区域的问题,提出了一种候选区域二次筛选法.结果表明:优化后,利用训练好的类Haar+Cascade Adaboost分类器、LBP+Cascade Adaboost分类器和HOG+Cascade Adaboost分类器对测试集马铃薯图像进行测试,其检出率、虚警率、总体识别率分别为1.7％、0.8％、97.2％;95.9％、0.0％、98.9％ 和86.7％、3.5％、93.9％;耗时分别为8.2、7.5和30.3 ms.这说明基于级联自适应提升(Cascade Adaptive Boosting,Cascade Adaboost)分类器的马铃薯定位方法,可快速准确定位运动中的马铃薯目标,其中LBP+Cascade Adaboost分类器的效果最优.