Vlib Experiment of DSP

Disclaimer: Reference Please specify source http://blog.csdn.net/lg1259156776/
Introduction

In DSP development, in order to save development time and difficulty, ti encapsulates some mature algorithms as modules for developers to use. If we can make full use of these algorithms to support the library, it is very helpful to speed up the DSP development process and improve the quality and stability of the code. The Vlib is a well-optimized library of computer vision algorithms that supports the C64X and c64x+ processor cores and provides the function interface as shown.
　　
The following is a summary of the test experiments performed on DSP6455:

Canny operator

Good detection: The algorithm is able to identify as many of the actual edges in the image as possible.
Good positioning: Identify the edges as close as possible to the actual edges in the actual image.
Minimum response: Edges in an image can only be identified once, and image noise that may exist should not be identified as an edge.
In the image edge detection, the suppression noise and the edge accurate localization is unable to satisfy simultaneously, the canny operator strives to seek the best compromise between the anti-noise and the precise localization.
The steps to detect image edges using the canny operator are as follows:

1. Gaussian Image Smoothing
   

2. Calculate gradient amplitude and direction

3. Non-maximal value suppression

hysteresis threshold algorithm (dual threshold algorithm)

In my understanding, summed up as: to a pair of graphs, the first Gaussian smoothing filter, and then calculate the gradient, the non-maximal value suppression method to remove the false boundary, and then use two threshold-high threshold value and low threshold value of the non-maximal value of the image to be processed, to obtain two results. The high threshold value image is designed to ensure that the edges are as accurate as possible and at the expense of some marginal point loss. The low threshold image is designed to compensate as much as possible for the drop in the actual edge point of the high threshold value. Finally, the edge of the high-threshold image is joined to the contour, and when it reaches the contour short point, the low-threshold image in the 8 neighborhood of the point position can be connected to the edge of the contour. Repeat this step continuously until you connect image 2 together. The general high threshold is about twice times the low threshold value.

Hough Line Detection

The basic algorithm of detecting the set graph in image processing in Hough transform, it can detect the regular image of line and ellipse in image, and the generalized Hough Transform can detect irregular boundary.
Collinear points in a Cartesian coordinate system correspond to the collinear lines of the parametric space .

Morphological corrosion and swelling

First corrosion after expansion known as on (open)
In general, open operations can remove isolated dots, burrs, and small bridges (that is, dots that connect two areas), while the overall position and shape are constant. This is the function of the open operation.

First expansion after corrosion called closed (close)
In general, closed operations can fill small lakes (i.e., small holes), bridging small cracks, while the overall position and shape are constant. This is the function of the closed operation.

Unicom area Mark

The Unicom area is intuitively understood as an independent object in the image, and the distribution of these unicom regions is marked by different labels, which is the purpose of the China Unicom region. In addition, these areas can be described using their internal features, such as area area and regional center of gravity.
Algorithm idea: Similar to the regional growth method, take 8 neighborhood Unicom region Mark as an example, to scan the image, find a non-background pixel, assign an unused tag, to the Point 8 neighborhood of the non-background pixels allocated with the same marker, next to these pixels in the 8 neighborhood of the non-background pixels to do the same operation, Iterate over and over until the 8 neighbors of these points can no longer find non-background pixels. Then proceed to scan, find a new unmarked non-background pixel, repeat the above operation, until you can no longer find any unmarked non-background pixels.

Commonly used two times scanning method

Color space

Mainly in the video processing or color tracking encounter color space conversion, color space including HSI, HSV, RGB, CMYK, YUV, lab and so on. The most common is the RGB space, from RGB24 to HSI conversion formula:

Integral image

Integral image is a kind of middle representation of image which is used to calculate the sum of gray value or gray mean of image window quickly, and it can be used to calculate the gray variance of image window quickly, and to realize the standardization of fast grayscale distribution in image preprocessing and the fast template matching in target task recognition. is an important intermediate step in calculating image features. There are many applications in face detection.
There are more and more familiar, not to repeat.

Image Pyramid

Image pyramid is a data structure, image pyramid decomposition can be very good to achieve the image of the foreground and background of the segmentation, and even can be used to contain multiple prospects of complex images. Select from the original image can represent a small area of the point as a node, so that these nodes constitute a higher layer of the image, and continue to iterate, the node of the graph is gradually reduced from top to bottom, the formation of pyramid shape, and wavelet decomposition similar, pyramid is a fine to rough strategy. It is mainly applied to different scales of image or target.
This has been summed up in the study of the optical flow method, there is no more to say.

Gauss and Gradient pyramids

Gaussian pyramid is a kind of image pyramid. Its No. 0 level contains the original image, the first level contains the 2x2 under-sampled images, the second stage contains further 2x2 under the sampled image, and so on. are often used in detection and tracking applications to reduce redundant data.
For image decomposition and reconstruction is the Laplace pyramid, Gaussian pyramid decomposition is a step in the decomposition of the Laplace pyramid.
Contourlet decomposition process can be expressed as: low-pass filtering → down sampling (reduced size) → interpolation (enlarged size) → bandpass filter (image subtraction).

Reconstruction of Contourlet

This is a study of the optical flow method, so you can check it again.

IIR Filter

IIR must adopt recursive structure, the pole in the unit circle, or the system is unstable. IIR filter is widely used in a series of image processing operations, such as smoothing, gradient and edge calculation.
No longer detailed table!

Angular point detection

Harris Corner Point Detection. This should be more familiar. For Lucas Harris, the light flow method seems to be using this corner!

Non-maximal value suppression

It has been used in the experiment of canny operator.
Some video algorithms result in a polling space, and the search for the local maximum in the polling space can be achieved with a non-maximal value suppression.

Optical Flow method

The most popular Lucas-kanade algorithm: the LK algorithm.

Grayscale Histogram statistics

Histograms are often used to measure the distribution of input data, and by using weighted histograms, it is possible to determine the importance of certain quantities in the input data. The histogram calculation is very simple, first by the specified calculation range, and divides the range into several adjacent units (bin), and then the number of input data falling into each adjacent cell is counted.

L1 distance and B distance

In mathematics and engineering applications will use a variety of distances, L1 distance, L2 distance, Chebyshev distance, MA distance, and so on, these distances to reflect the difference between different things, such as points, polygons, probability distribution.
L1 distance becomes also block distance, Manhattan distance or absolute distance, L1 norm.

b ha tt a c h a ry y a distance, also known as Babbitt distance or B distance, in statistics, used to measure the similarity of two discrete probability distributions, under normal conditions, the Babbitt distance is used to measure the separation between the two categories.

Kalman Filter

This is quite a lot in the previous application, so the specific principle does not have to be analyzed. For how to use in the DSP, you can follow the specific use of the changes in accordance with the routine.

Weighted moving average and weighted moving variance

The application of weighted moving average and weighted moving variance is very wide, and can be used in the field of moving target detection in digital image processing. In the background extraction process of moving target detection, it is necessary to update the background information in time due to the influence of illumination and other factors. To remove the effect of the foreground object on the background, updates to the weighted moving average and the weighted move variance occur only in the Background section, where the foreground area is marked as 0.

Static background extraction

Used to split the background and foreground of a static image. If the background is static, then the moving object with any meaning is the foreground.

Gaussian mixed background model

The central content of intelligent monitoring is moving target detection, the key is the extraction of background objects, the purpose of the modeling background is to extract the foreground from the current frame, make the background closer to the background of the current frame, Gaussian mixture model is one of the best modeling effects in many models.
Gaussian mixed model idea: For each pixel, to examine its mean and variance, if the match with a Gaussian model (the deviation is less than the Markov distance threshold), then the point is divided into this model, while the model of the three parameters: mean variance and weights are to be updated, if there is no Gaussian model to match the point, The most unlikely of these Gaussian models (the one with the lowest exit rate), and a new Gaussian model is established at that point.

2015-9-23 Less Art

Copyright NOTICE: This article for Bo Master original article, reprint please indicate source http://blog.csdn.net/lg1259156776/.

Vlib Experiment of DSP