Paul Cockshott Images In Java ALMA TADEMA RIVALES INCONSCIENTES 1893 Introduction • The AWT and how to put a simple image onto the screen • Layout managers • The Jimage Class, Summary of this section • At the end of this lecture you should have an idea of how to display a JPEG image on the screen, and how to load it into the Jimage class to carry out further image processing. Agenda • • • • • AWT Images Image Producers and Consumers Jimage class Pixel Representations JPEG files Overview • AWT abstract windows toolkit, supported by JavaSoft • Operating system independent layer for windowing in Java • Fiendishly obscure • Designed around requirements of images being streamed off the web Connections • Simple image display program to show how to display a JPEG file • Pipeline model of image production • Jimages act as image consumers • Jimages allow arithmetic on image • Jimages provide output to AWT images and JPEG How to display a picture 1 import java.awt.*; import java.awt.image.*; import java.util.*; class JPEGshow extends Frame { ... static public void main(String[] args) { if (args.length == 1) new JPEGshow(args[0]); else System.err.println("usage: java JPEGshow <image file>"); } } This is a standard Java Program class with a public static void main method Constructor for JPEGshow See slide on these JPEGshow(String filename) { super("JPEG show Example"); add( new ImageCanvas(getToolkit().getImage(filename) ), BorderLayout.CENTER); setSize(700, 540); show(); } Read in a JPEG file Size of a frame The toolkit • Each frame has associated with it a toolkit object the provides an interface to OS specific operations. • • • • CreateImage CreateMenu CreateLabel CreateMenuBar …. etc Roll your own ImageCanvas Constructor class ImageCanvas extends Component { the image Image image; ImageCanvas(Image image) {this.image = image;} public void paint(Graphics g) { g.drawImage(image, 0, 0, this);} } just stores Paint is called whenever a component must be shown, the Graphics object does the actual drawing, it has to be passed in because it is what knows about physically drawing on the screen Image Class • Pipeline flow model of image processing • Images are just tokens linking producers and consumers ImageProducer Image ImageConsumer ImageProducer Methods • • • addConsumer(ImageConsumer ic) This method is used to register an ImageConsumer with the ImageProducer for access to the image data during a later reconstruction of the Image. removeConsumer(ImageConsumer ic) This method removes the given ImageConsumer object from the list of consumers currently registered to receive image data. startProduction(ImageConsumer ic) This method starts an immediate reconstruction of the image data ImageConsumer methods • void setDimensions(int width, int height) The dimensions of the source image are reported using the setDimensions method call. • Void setPixels(int x, int y, int w, int h, ColorModel model, byte[] pixels, int off, int scansize) The pixels of the image are delivered using one or more calls to the setPixels method. Image Class continued ImageProducer Image ImageConsumer Image.getSource Images contain a pointer to their producer which holds the actual data for the image. This can be recovered using the getSource method. This allows a consumer to get at the pixel data of an image by adding itself to the producer and starting production Summary • AWT is operating system independent • Streaming image model • Images as tokens • Producer - consumer pipeline • See chapters 6 of textbook Buffered Image Class • Standard AWT images are just tokens for data streams. • A BufferedImage actually contains the data. BufferedImage Colour model Raster JPEGCodec class • This class has factory methods to create JPEG encoders and decoders: • createJPEGDecoder(InputStream s) • createJPEGEncoder(OutputStream d) Read a BufferedImage FileInputStream in = new FileInputStream(“myfile.jpg”); JPEGImageDecoder dec= JPEGCodec.createJPEGDecoder(in); BufferedImage im = decoder.decodeAsBufferedImage(); getRGB • You can access the pixels of a buffered image using int getRGB(int x, int y) The default colour representation is: alpha Bit 31 red green blue Bit 0 Writing pixels • This can be done with the setRGB method. • This takes x, and y co-ordinates and a pixel encoded as a 32 bit integer • im . setRGB(2, 5, 255); • Would set pixel 2,5 to 255 = bright blue. Creating sub images • You can create a sub area within a buffered image using the • public BufferedImage getSubimage( – – – – int int int int x, y, w, h); Method of BufferedImage Jimage implements ImageConsumer • Library of image processing classes developed in the department • Available for student practicals • Algebraic rather than stream oriented • Interfaces to MMX hardware under windows Algebraic orientation By this we mean the it is structured around algebraic expressions whose values are images Thus if A and B are images and is some operator then AB is also an image Jimage operators • • • • • • • Arithmetic I+J Universal plus(Universal) I-J Universal minus(Universal) I×J Universal times(Universal) I÷J Universal divide(Universal) IUniversal abs() Filtering Jimage convolve(double[] k) convolve with symmetrical separable kernel. public abstract Jimage convolve(double[][] kernel)with non separable kernel Scaling • Jimage getScaledInstance(int nwidth, int nheight) This scales with bicubic interpolation. • Jimage getScaledInstance(int nwidth, int nheight, int ndepth) This method allows the depth as well as area of an image to be altered if it is reduced the planes are aggregated if increased they are interpolated. More operations Data access int rgbpixel(int x,int y) Converts the plane information into a pixel in the direct color model of java. public abstract int upixel(int x, int y, int plane) - returns unsigned integer pixel public abstract float fpixel(int x, int y, int plane) Returns the pixel in the range -1 to +1. Data Access • • • • public abstract void setPixel(int x, int y, int plane, double pix) • • • • – Pixel information in range -1 to +1 public void setSubImage(int x, int y, int z, Jimage im) – Update an area of an image with another one. The other one must not run off the edge of the one being written to. The source of the copying is the 0th plane of the source jimage. Jimage input output • public void putJPEGImage( • java.lang.String fileName, • int quality) • throws java.io.IOException – Outputs the image to a jpeg file • public boolean getImage(java.lang.String fileName) – Initialise the Jimage from the specified file. The file must be jpeg or gif. Jimage to AWT Image conversion • public java.awt.Image getAWTImage() • public java.awt.image.ImageProducer getProducer() Jimage implementations JIMAGE CLASS HIERARCHY COM.C3D.IMAGE Jimage abstract COM.C3D.IMAGE ByteImage Generic Java COM.C3D.IMAGE ShortImage Generic Java COM.C3D.IMAGE FloatImage Generic Java IntelBImage Runs best on MMX Windows only IntelImage Runs best on MMX Windows Only IntelFImage Runs best on PIII Windows Only An example program • • • • • • • • • • • class Jimageshow extends Frame { Create Jimage Jimageshow(String filename) { with byte pixels super("Jimage show Example"); Jimage raw=new ByteImage(100,200,3); if (raw.getImage(filename)){ Jimage cooked = (Jimage)raw.times(0.3); Multiply by 0.3 add(new ImageCanvas(cooked.getAWTImage()), BorderLayout.CENTER); setSize(700, 540); show(); } } Convert to AWT for display Pixel Representations When dealing with displays it is conventional to assume that pixels are bytes holding numbers in the range 0 to 255. 0 Is assumed to be black 1 Is assumed to be white or maximum brightness of any given colour. For multicolour displays with 3 colour components, the convention is to have 3 fields of range 0..255 to hold the colour information. Pixel Representations AWT For multicolour displays with 3 colour components, the convention is to have 3 fields of range 0..255 to hold the colour information. The AWT does this with the class Color. • public Color(int rgb) – Creates a color with the specified RGB value, where the red component is in bits 16-23 of the argument, the green component is in bits 8-15 of the argument, and the blue component is in bits 0-7. The value zero indicates no contribution from the primary color component. – A Jimage returns this format with int rgbpixel(). Pixel Representations: Bytes The byte data type in Java does not take on the values 0..255. Instead it takes on the values -128 to 127. There are no unsigned bytes in Java. This creates a problem for the representation of pixels in Jimages. The solution adopted is to adopt the following representation • -128 = black • 0 = mid grey • 127 = white Pixel Representations: Floats If byte pixels are signed then so must other representations be. The solution adopted is to adopt the following representation for floats • -1 = black • 0 = mid grey • 1 = white Conversions between representations unsigned min value -1 maxval 255 medianval m 0 range r 255 bytes shorts 0 -128 -2048 127 2047 127.5 -0.5 255 4095 float 1 -0.5 2 As shown in table a pixel prin representation r is converted to a pixel ps in representation s by the operation: ps = ms+(rs(pr-mrrr Signed Pixels : advantages Signed pixels seem at first to be counter-intuitive but they have numerous advantages. • • • A value of 0 or mid grey can be viewed as the ‘most likely’ value that a pixel takes on in the absence of other information. If you do arithmetic on images, in particular subtract one image from another, then negative values of pixels naturally arise. Signed pixels allow straightforward implementation of contrast adjustments. For instance multiplying an image by 0.5 halves the contrast in the image. Signed Pixels : contrast adjustment Signed pixels allow straightforward implementation of contrast adjustments. For instance multiplying an image by 0.5 halves the contrast in the image. 0.5 0.5 1 0.5 0.25 -0.25 -0.5 -1 Initial contrast range Finalcontrast range Image Multiplication Image Addition + Image subtraction - What Is Convolution • Convolution takes a kernel of coefficients and multiplies each pixel in a neighbourhood by the corresponding coefficient, and then sums the result • x y p[I+x, j+y]*k[x,y] • Will give the convolved pixel at position i, j 1 D convolution • A 1 D convolution takes a one dimensional array as a kernel and applies it first in the X and then in Y dimension. • This can often be performed faster than a 2d convolution Image Convolution: smoothing marble = double[] k= Note sum of coefficients =1 marble.convolve(k)= {0.1,0.1,0.2,0.2,0.2,0.1,0.1}; Image Convolution: sharpening marble = double[] k= Note sum of coefficients =1 number terms is odd marble.convolve(k)= {-0.3,1.6,-0.3} Convolution in Java2D • Java 2D provides a standard library for convolution of buffered images • This uses the class Kernel and ConvolveOp Kernels in JAVA 2D float[] blur={ 0.0f, 0.1f, 0.0f, 0.1f, 0.6f, 0.1f, 0.0f, 0.1f, 0.1f}; Kernel k= new Kernel(3,3, blur); im = new ConvolveOp(K).filter(im,null); This will blur the image im by applying the 3 by 3 kernel blur to it. Importance of speed • Image may contain a million pixels, • Arithmetic may be required on each one • Important to optimise operations or they are very time consuming • May need to use assembler kernels • May need to use special purpose instructions Multimedia Extensions MMX • Intel and other CPU manufacturers have been adding to the instruction sets of their computers new extensions that handle multi-media data. • The aim is to allow operations to proceed on multiple pixels each clock cycle MMX 2 • Standard Intel register set 8 General Registers 8 floating point registers fp0 fp1 fp2 fp3 fp4 fp5 fp6 fp7 eax ebx ecx edx esp ebp esi edi 32 bit 64 bit MMX 3 • Standard Intel register set operating in MMX mode 8 General Registers 8 multimedia registers mm0 mm1 mm2 mm3 mm4 mm5 mm6 mm7 eax ebx ecx edx esp ebp esi edi 32 bit 64 bit MMX 4 motivation • Existing operating systems must still work unchanged • Applications not using MMX run unchanged • No new state added to the CPU Hence, shared use of the FP registers, since these are already supported by exising OS’s MMX data formats One 64bit integer QUADWORD Two 32 bit integer DOUBLEWORDS Four 16 bit WORDS Eight 8 bit BYTES Problem of overflows • • A problem with limited precision arithmetic is that overflows frequently occur. This can give rise to meaningless results: consider 200+175 = 375 but in 8 bit binary 11001000 +10101111 =101110111 Leading 1 is discarded This leaves an answer of 119 decimal – clearly wrong Using saturation • You can fix this by using conditionals unsigned char p1,p2,p3; int I3= (int)p1 + (int)p2; p3=(I3>255?255:(unsigned char)I3); Expansion of the code 1 12: 00401043 00401046 00401048 0040104A 0040104D 0040104F 00401051 00401053 00401056 00401059 0040105C 0040105F 00401062 00401065 mov xor mov mov xor mov add mov mov add mov mov add mov j=(int)(*p1++)+(int)(*p2++); ecx,dword ptr [ebp-4] edx,edx dl,byte ptr [ecx] eax,dword ptr [ebp-8] ecx,ecx cl,byte ptr [eax] edx,ecx dword ptr [ebp-14h],edx edx,dword ptr [ebp-8] edx,1 dword ptr [ebp-8],edx eax,dword ptr [ebp-4] eax,1 dword ptr [ebp-4],eax Expansion 2 13: 14: 00401068 0040106F 00401071 00401078 0040107A 0040107D 00401080 00401083 00401086 15: 00401088 0040108B 0040108E *p3 = (unsignedchar)(j>255?255:j); cmp dword ptr [ebp-14h],0FFh jle main+6Ah (0040107a) mov dword ptr [ebp-18h],0FFh jmp main+70h (00401080) mov ecx,dword ptr [ebp-14h] mov dword ptr [ebp-18h],ecx mov edx,dword ptr [ebp-0Ch] mov al,byte ptr [ebp-18h] mov byte ptr [edx],al p3++; mov ecx,dword ptr [ebp-0Ch] add ecx,1 mov dword ptr [ebp-0Ch],ecx Total of 26 instructions in the kernel Alternative using mmx Type represents 8 by 8bit integers • • • • • • • Iu8vec8 *v1,*v2,*v3; int i,j,k; for(i=0;i<31;i++){ *v3=(*v1++)+(*v2++); v3++; } _mm_empty(); Arithmetic on 8 bytes at a time Indicates MMX regs are now free Optimised Assembler Loop Go round only 32 times not 256 mov ecx ,32 ; load counter with 32 l1: movq mm0,[esi] ; load 8 bytes add esi,8 ; inc src pntr paddusb mm0,[edx] ; packed unsigned add bytes add edx,8 ; inc src pntr movq [edi],mm0 ; store 8 byte result add edi,8 ; inc dest pntr loop nz,l1 ; dec counter, ; repeat non zero Total of 6 instructions in kernel Speed Gain • On image of 256x256 pixels • Old C code executes 26*256*256 instructions = 1,703,936 instructions • Optimised mmx code executes 6*256*32 instructions = 49,152 • Note that no compiler currently will give the optimised code. It has to be hand assembled. Image Processing Library • Intel Provide an image porcessing library that can be downloaded from their web site. • It provides efficient access to the MMX hardware. • It provides frequently used Image Processing Operations. • It requires a set of DLLs in your path to run Image Processing Library 2 At the core of IPL is the ability to write to a single API and get the best possible results for any Intel processor. The libraries have as many as six processorspecific branches for each function and six sets of carefully written assembly code, but only one entry point to each function. Image Processing Library 3 Image Processing Library 4 • • • • • Use of Intel IPL complex and requires C I have provided 2 java classes that call the IPL. IntelBImage and IntelFImage. These are documented in the Jimage web pages. They inherit from ByteImage and FloatImage To use them the Intel IPL must have been installed on your machine and be on the path. If you are forced to use Unix machines the libraries will not be available to you. Where to get more information • http://www.javasoft.com/products/jdk/1.2/d ocs/api/java/awt/package-summary.html • http://developer.intel.com/vtune/perflibst/ip l/index.htm • http://developer.intel.com/vtune/perflibst/ip l/ipapi.htm Feedback • What did you not understand here? • What would you like more information on?