基于softcascade的车牌检测算法

Softcascade-based Plate License Detection Algorithm

相较于传统的分类器级联方式,softcascade分类器具有很多优点：softcascade中每一级的输出为之前所有弱分类器输出值之和,这样前面选出的特征会参与后面每一级的决策,得到分类性能更好的分类器。采用softcade算法训练分类器,结合扩展的HARR特征实现了车牌检测,并引入并行计算,在GPU设备上对整个训练过程进行了优化,使训练速度提升了3~4倍。与传统级联算法进行对比实验,实验结果表明基于softcascade算法训练得到的分类器较传统的级联算法训练得到的分类器的性能具有较大提升。

Compared with traditional classifier cascade structure,softcascade classifier structure has many advantages.Softcascade structure tend to get better classifier as the features choose in early stages will be used to make decision in all later stages.In this paper,a plate License detector is realized using extended HARR features and softcascade algorithm,the training process is optimized on the GPU device and the training speed improved three to four times.