Method for training multi-genus Boosting categorizer

Abstract

The invention discloses a method for training multiclass Boosting categorizers. The method is characterized in that a class weight corresponding to the class attributed to a training sample is distributed for a sample weight of each training sample in the training data after each circulation and before starting the next circulation in the training process, that is, each training weight of the training sample in the training circulation comprises the sample weight and the class weight. The class weight corresponding to each class obtains a strong categorizer according to the circulation training and the latest circulation training towards the performance and the dynamic change of the class, so that the training weight of the training sample of the class with poor performance in the next circulation is enlarged, the training weight of the class with good performance in the next circulation is reduced, and the performance of each class reaches the target threshold of the performance as far as possible in the same circulation to achieve the training. Therefore, the invention eventually ensures that the quantity of the weak categorizer required by the class with the worst performance is reduced; and meanwhile, the quantity of the weak categorizer required for categorizing different classes is basically the same.

Description

technical field [0001] The present invention relates to a classifier training method, more specifically, relates to a method for training a multi-class Boosting classifier by dynamically changing the class weights and adjusting the training weights of training samples when using the multi-class Boosting method for class training. Background technique [0002] Multi-class boosting methods are very important for multi-class object detection and recognition, especially in the field of image object detection. Object detection in images is becoming more and more popular and is used in several different detection classes. Examples of its applications include: multi-view face detection, eye location, traffic signal detection, and vehicle detection. [0003] figure 1 A multi-class Boosting method for multi-view face detection is shown. According to the rotation angle of the face relative to the image plane, multi-view faces are divided into several categories. The multi-class Bo...

AI Technical Summary

Problems solved by technology

This makes the whole system slower and less efficient

Images