Programming Assignment 2 CS308 Fall 2001

Goals Improve your skills with using templates. Learn how to compile your code when using templates. Learn more about image processing. Learn to document and describe your programs.

Template the Image class Make the implementation general enough so it can handle both gray-level (PGM) and color (PPM) images. Download the code for reading/writing PPM image from the course’s webpage. template class Image { public: member functions … private: int N; // no of rows int M; // no of columns int Q; // no gray-levels PixelType **pixelVal; }; PixelType can be int or RGB

New class for color pixels Each color pixel is represented by three values: (r,g,b) struct RGB { int r; // red component int g; // green component int b; // blue component } ;

Important things to remember when using templates Your implementation should be as general as possible. Your code should not depend on gray-level/color images. If you have to use the Image class with a new type of images in the future, then you shouldn’t have to make any changes to the implementation of the class ! If you need to make changes, then your are not hiding the implementation details from the application ! This is bad programming strategy.

An example (without using templates) int Image::meanGray() { int i, j; int avg; avg = 0; for(I=0; I<N; I++) for(j=0; j<M; j++) avg = avg + pixelVal[I][j]; avg = avg / (N*M); return avg; }

An example (using templates) template PixelType Image::meanValue() { int i, j; PixelType avg; avg = 0; for(I=0; I<N; I++) for(j=0; j<M; j++) avg = avg + pixelVal[I][j]; avg = avg / (N*M); return avg; }

An example (cont’d) How does the function know how to initialize a variable of type PixelType to 0? Hoes does the function know how to add together two variables of type PixelType? How does it know how to implement assignment for PixelType variables? How does it know how to divide a variable of type PixelType by an integer?

An example (cont’d) Actually, it doesn’t until the user of the class has defined PixelType (i.e., by passing a parameter to the template) Image myGrayLevelImage; Image yourColorImage; Where does the function look for the definition of these operations? “You need to define these operations within the class RGB”

Specification and Implementation Separating the specification from implementation makes it easier to modify programs. Changes in the class’s implementation should not affect the client (as long as the class’s interface has not been changed). Follow the following two principles: –Place the class declaration in a header file to be included by any client that wants to use the class (e.g., StackType.h). –Place the definition of the class member functions in a source file (e.g., StackType.cpp)

Multiple definitions When splitting a program into multiple files, the header file needs to be included in each file in which the class is used. To avoid multiple definitions of a class (or even functions), the class specification need to be enclosed in the following preprocessor code: #ifndef NAME #define NAME #endif

Multiple definitions (cont’d) #ifndef STACKTYPE_H #define STACKTYPE_H template class StackType { public: …….. void Push(ItemType); void Pop(ItemType&); private: int top; ItemType *items; }; #endif

Rules when using templates When working with templates, we change the ground rules regarding which file(s) we put the source code into. Previously (no templates) StackType.cpp could be compiled into object code independently of any client code. Using templates, the compiler cannot instantiate a function template unless it knows the argument to the template This information is found in the client code !!

Rules using templates (cont’d) Two possible solutions –Place the class definition and member function definitions into the same file (e.g., StatckType.h) –Or, give the include directive for the implementation file at the end of the header file

Rules when using templates (cont’d) template class StackType { public: …….. void Push(ItemType); void Pop(ItemType&); private: int top; ItemType *items; }; #include StatckType.cpp

Thresholding It computes a binary (black/white) image of the input.

threshold(image, thresh) Implement this as a client function (needs to be overloaded for int and RGB). Each pixel in the input image is compared against a threshold. Values greater than the threshold are set to 255, while values less than the threshold are set to 0. The same idea applies to color images (using up to three thresholds now)

Coin segmentation Separate the regions corresponding to the coins from the background. Useful for coin recognition: –collect all the pixels belonging to the same region –extract “features” useful for coin recognition

Character segmentation original thresholoded

Object counting Use up to three thresholds in this case: –a threshold for red –a threshold for green –a threshold for blue

Face segmentation originalthresholded candidate face regions

Important Issues Regarding Thresholding How to choose the threshold? How to improve the results of thresholding? How to segment specific objects only?

How to choose the threshold? originalgood threshold low threshold high threshold

displayHistogram(image) Implement this as a client function (needs to be overloaded for int and RGB). The histogram is a bar graph of the pixel value frequencies (i.e., the number of times each value occurs in the image)

displayHistogram(image) -- cont’d Use an array of counters to store the pixel frequencies. Display the histogram as another image. Draw a bar for every counter. Normalize counter values:

Face segmentation Find regions containing skin-color. In practice, we would have to estimate the skin-color distribution using statistical models and lots of images containing regions of skin-color. You will try a simpler approach here: –convert RGB values to HSV values (Hue-Saturation-Value) –threshold the HSV values

RGBtoHSV(RGB) Converts a pixel from RGB to HSV

faceThreshold(image) You should implement this as a separate client function since it is specific to face detection. This function should call RGBtoHSV first. Then, it will apply the thresholding shown below to compute a binary image (we do not use V):

How to improve the results of thresholding? In most cases, further processing is required to improve the results of thresholding. For example, some of the regions in the thresholded image might contain holes.

dilate(image) -- client If at least one neighbor is 255 if all 8 neighbors are 0

dilate(cont’d) Dilation “expands” the regions (i.e.,adds a layer of boundary pixels) originalthresholdeddilated

erode(image) -- client If all 8 neighbors are 255 if at least one neighbor is 0

erode(image) Erosion “shrinks” the regions (i.e., removes a layer of boundary pixels) originalthresholdederoded

Filling in the holes of regions Apply dilation to fill in the holes. Apply erosion to restore the size of the regions. originalthresholded dilated eroded

How to segment specific objects only? There are cases where we are interested in detecting some of the objects in the scene (e.g., just the key and the coin). Can you do this using the simple threshold function?