Normalization of images

First, Introduction

Image normalization is a kind of technology widely used in computer vision, pattern recognition and other fields. The so-called image normalization, is through a series of transformations, the raw image to be processed into the corresponding unique standard form (the standard form of image for translation, rotation, scaling and other affine transformations have invariant characteristics).  In recent years, the technology of image normalization based on moment has been paid much attention to, and its basic working principle is: firstly, using the invariant moment of affine transform in image to determine the parameters of the transformation function, and then using the transformation function determined by this parameter to transform the original image into a standard form image (the image is independent of affine transformation). In general, the process of image normalization based on moments consists of 4 steps: Coordinate centering, x-shearing normalization, scaling normalization and rotational normalization.

Basically, the idea of normalization is to use the invariant moment of the image to find a set of parameters so that it can eliminate the influence of other transformation functions on the image transformation. That is, to convert to the only standard form to resist affine transformations. The image normalization makes the image resistant to geometric transformations, and it can find the invariants in the image so that the images are the same or a series. Here's what you need to know: 1. Normalization does not change the contrast of the image 2. Normalization processing is very simple, assuming the original image is a 8-bit grayscale image, then the reading of the pixel matrix maximum value is 256, the minimum value is 1, the definition of the matrix is i,j=i/256, is the normalized image matrix, This means that all pixel values are within the [0,1] interval after normalization.

Second, what is normalization

Normalization is through a series of transformations (that is, using the invariant moment of the image to find a set of parameters so that it can eliminate the effect of other transformation functions on the image transformation), the original image to be processed into a corresponding unique standard form (the standard form of image for translation, rotation, scaling and other affine transformations have invariant characteristics).

The basic working principle of image normalization based on moment is: firstly, the parameters of transform function are determined by using the invariant moment of affine transform in the image, and then the transformation function determined by this parameter transforms the original image into a standard form (the image is independent of affine transformation). In general, the process of image normalization based on moments consists of 4 steps, i.e., coordinate centering, x-shearing normalization, scaling normalization and rotational normalization.

The image normalization makes the image resistant to geometric transformations, and it can find the invariants in the image so that the images are the same or a series.

Third, why normalization

1. Basically the idea of normalization is to use the invariant moment of the image to find a set of parameters so that it can eliminate the influence of other transformation functions on the image transformation. That is, to convert to the only standard form to resist affine transformations. The image normalization makes the image resistant to geometric transformations, and it can find the invariants in the image so that the images are the same or a series.

2.matlab image data must sometimes be floating-point type to process, and the image data itself is 0-255 of the unit type data, so need to be normalized, converted to 0-1.

3. Normalization is a way of simplifying computation, and the expression of dimension will be transformed into a dimensionless expression to become a pure quantity. The purpose is to:

(1) Avoid the unequal use of input variables with different physical meaning and dimension

(2) sigmoid function is often used as transfer function in BP, normalization can prevent neuron output saturation caused by large net input absolute value.

(3) To ensure that the output data of small values are not swallowed

3. Reasons for normalization in neural networks:

Normalization is to speed up the convergence of training network, can not be normalized processing;

The specific function of normalization is to summarize the statistical distribution of uniform samples. Normalization between 0-1 is the probability distribution of statistics, and normalization between -1--+1 is a statistical coordinate distribution. Normalization has the same meaning, unity and oneness. Whether it is for modeling or for calculation, first of all the basic units of measurement to the same, the neural network is the sample in the event of statistical probability of training (probability calculation) and prediction, normalization is the same in the 0-1 of the statistical distribution; When the input signal of all samples is positive, The weights associated with the first hidden layer neurons can only be increased or decreased at the same time, resulting in a slow learning rate. In order to avoid this situation, speed up the network learning speed, the input signal can be normalized, so that all samples of the input signal its average value is close to 0 or less than its mean variance.

Normalization is because the value of the sigmoid function is between 0 and 1, the output of the last node of the network is the same, so often the output of the sample is normalized. So it is better to use [0.9 0.1 0.1] to classify problems than [1 0 0].

But normalization is not always appropriate, and other statistical transformations, such as normalization, can sometimes be better based on the distribution of output values.

from:http://blog.csdn.net/u010555688/article/details/25551255

Normalization of images