改进感知机多类分类方法在车辆类型识别中的应用

APPLICATION OF IMPROVED PERCEPTRON MULTI-CLASS CLASSIFICATION METHOD IN RECOGNITION OF AUTOMOBILE MODELS

车辆类型识别特征参数提取的关键是确定车辆顶棚位置。提出一种按比例截取顶棚厚度、确定顶棚位置的方法。针对训练传统感知机所需初始权向量及惩罚系数需人工设定而造成训练速度慢、准确率低等问题,提出一种初始权向量自动确定及惩罚系数更新方法,并应用一对一分类方法解决车型繁多分类问题。实验结果表明,该方法能有效收敛到固定权值,顶棚位置确定准确,迭代速度快,识别准确率高。

The key of characteristic parameters extraction for vehicle models＇ recognition is to determine the position of vehicle roof. We present a method to intercept the roof thickness proportionally for determining the position of a roof. Aiming at the problems of traditional per- ceptron that the initial weight vector and penalty coefficients it required have to be set manually but will cause slow training speed and low ac- curacy, we put forward a method which can automatically determine the initial weight vector and update the penalty coefficients. We also ap- ply the one-to-one classification method to solve the problem of various vehicle models classification. Experimental results show that our meth- od can converge to the fixed weight vector effectively, the position of roof is determined accurately, the iteration speed is fast, and the method has high recognition accuracy as well.