Inline Cascade classifier

Inline cascade classifier nested CASCADE detector detector AdaBoost Real AdaBoost Read "C. Huang, H. Ai, B. Wu, and S. Lao, ' boosting Nested Cascade Detector for Multi-View face Detection ', ICPR, 2004,vol ii:4 15-418 "notes

Main contribution points of thesis

• This paper presents a weak classifier based on Haar feature lookup table, and uses real adaboost to learn a strong classifier.

• A nested cascade classifier (nested CASCADE detector) is proposed, which can improve the classification performance better and faster. By the introduction of nested weak classifier, the number of features is greatly reduced.

Real AdaBoost Discussions

In fact, this paper does not improve the real adaboost algorithm, but the cascade detector in the classifier design adaboost replaced the real AdaBoost method.

The main contribution point of this aspect, I think the body now, the article from the angle of the Bhattacharyya Distance to explain the real adaboost in the process of feature selection reasons.

We have already explained that when Real AdaBoost chooses a feature, it divides a certain dimension feature space into several intervals and then classifies it in different intervals, and the output of its classifier is

and the choice of feature is

which

We deduced from the minimization of the exponential error, and now we understand the process of feature selection from the angle of Babbitt's distance.

For any two random distributions, the Babbitt distance is defined as:

In the search for features, it is expected that the distribution distance of positive and negative samples is as large as possible, so it is easier to distinguish positive and negative.
When the real adaboost is used, each dimension is divided into n intervals, then the bar distance of the positive and negative sample distributions on the dimension is:

So we are maximizing when we are minimizing.

In conclusion, we state this in order to train a best lut-type weak classifier under a distribution $D _t$, first a Haar FE Ature which gets the maximum $J _b$ should be selected.

Here Lut-type refers to the look-up-table type, because each feature is divided into n intervals, each of which has corresponding output, so it can be categorized by a lookup table.

Nesting cascading classifiers
The training sample for the next-level classifier in the Cascade classifier is a sample checked out by the previous classifier, and if the classifier is trained using real adaboost, then the result of the classifier output is not just a positive and negative sample of the function, but a definite real value, simply will ' Positive samples ' equal treatment of input to the next level of the classifier, greatly ignoring the real value of the output of the difference.

Each classifier in the Cascade classifier requires a higher detection rate, and the false detection rate is within acceptable range, which causes the classifier to detect a positive sample that still has a noticeable degree of scalability.

As shown in the distribution, the left image is the original distribution of the sample, and after the first level of the AdaBoost classifier, all the ' positive samples ' are distributed as shown in the image on the right. Easy to find in order to obtain a higher detection rate, there are many real negative samples in the ' positive sample ', and the distribution of these samples can be found to be better separated.

Resample.png
Therefore, in order to use superior classifier to get the classified information of real value result, on the basis of the original cascade detector, a weak classifier is added in each layer, and the characteristic of the weak classifier is the real value output given by the previous classifier.

On the whole, nested cascade detector is composed of four parts: Haar feature, general weak classifier, embedded weak classifier (using upper layer real value output classifier) and strong classifier, its structure diagram is as follows:

Nested.png

Nested weak classifier can be constructed using inter-partition methods, so that the output of real adaboost for each layer is:

The output of the final classifier is

K is the classifier layer number.

Note: In the design of each layer of strong classifier, the nested weak classifier is used as the first weak classifier.

Experimentally, conf (x) have a much larger bhattacharyya distance than a Haar feature, so the nested weak classifier is a ' Strong ' One, which is placed as the first component of each layer except for the first layer.

Multi-View Face detection
Using nested cascade detector to realize the face detection in multi-view, the face rotation-out-plane attitude is divided into 5 categories: Left face, left side face, positive face, right half face, right face, and then construct a nested for each class. Cascade Detector, experimental results show that this kind of application can achieve good results.

Is the detection process:

Mvfd.png
Here is a module, called Pose Estimator, which is equivalent to a pre-selection process, through the selection of postures, so that the sample can be relatively accurate into the corresponding classifier, speed up the classification.

Embedding cascading classifiers