Electron probe microanalysis EPMAWhat’s the point?Image Processing & AnalysisImage Enhancement - Done LaterIntensities, Histograms, LUTsBrightness and Contrast: or How I Learned to Love the HistogramGamma ProcessingHistogram Levels & EqualizationKernels/Rank OperatorsNeighborhood averagingImage Math2 Dimensional HistogramsProcessing in Frequency SpaceLook Up TablesProcessing binary imagesThresholdingMaking an Image into a BinaryBoolean* OperationsColor Superposition of Elemental MapsErosion/DilationImage measurementsResources:ConclusionProcessing Electron, X-ray, and CL imagesElectron probe microanalysisEPMA Modified 8/22/08What’s the point? “A picture is worth a thousand words”. Raw images sometimes need to be “processed” • to highlight particular features (sometimes easy to do, sometimes difficult and requires advances computation)• to extract quantitative information (e.g. modal abundances)Image Processing & Analysis• Image enhancement• Segmentation and thresholding• Processing in frequency space• Processing binary images• Image measurements• Image presentation• Histogram normalization: crunching from 16 to 8 bit. This usually is a first step for visual presentation purposes, as most software packages only operate on 8 bit images. However, this does not apply for measuring absolute values of pixel intensity, such as X-ray counts. • Brightness/contrast (and importantly, gamma): adjusting histogram “levels” • Histogram equalization: divide intensities into equal/weighted number of categories• Kernels/Rank operators: modify each pixel by some operation upon it and nearest neighbors• Image math: background subtraction; ratio 2 elements• Processing in frequency space (Fourier transform): removing periodic noise • Applying alternate lookup tables (LUTs) for improved presentationImage Enhancement - Done LaterAll images we are concerned with (e.g., BSE, CL, X-ray) contain one channel of information, where each constituent pixel has a value from 0 to 255 (28) or 65535 (216). These can be ordered in a histogram of intensities, with the spread defining the contrast, and the absolute values defining how bright or dark the image is. These INPUT intensities are mapped onto an OUTPUT grayscale or color table known as a Look Up Table (LUT). The transfer function is known as gamma. A gamma of 1.00 indicates a linear relationship between pixel intensities and grayscales. A gamma >1 is a non-linear function where the darker pixels are made preferentially brighter, whereas gamma <1 has the very bright pixels preferentially darkened somewhat. Adjusting only “brightness” and “contrast” controls (highlighted in many image packages) generally give poorer results compared to tweaking the gamma as part of histogram adjustment.LUTIntensities, Histograms, LUTsBrightness and Contrast: or How I Learned to Love the HistogramPhotoshopThe original histogram is too bunched up – poor contrast. Notice the top (input) left and right sliders are not close to the min/max brightness.So we move the top (input) left and right sliders in to the min/max brightness levels.And we move the bottom (output) sliders to 10 and 254.A last (important) step is to adjust the gamma, the top middle slider. To left (higher) increases brightness of mid grays (normally the best option).Adjust LevelsGamma ProcessingThe traditional imaging medium, photographic paper, has a non-linear response to light exposure through the overlying negative. Skilled darkroom technique used this to bring out subtle features in the shadows, or enhance bright features that tend to wash out. For digital images, such nonlinear processing, gamma processing, provides selective contrastGoldstein et al, 1992, Fig. 4.53, p;. 238enhancement at either the black or white end of the gray scale, while preventing saturation or clipping of the resulting image. The signal transfer function is defined as where  is an integer (1, 2, 3, 4) or a fraction (1/2, 1/3, 1/4) and K is a linear amplification constant. For =2, a small range of input signals at the dark end of the gray scale are distributed over a larger range of output gray levels, enhancing the contrast here; signals at the white end are compressed into fewer gray levels. For  =1/2, expansion occurs at the bright end, enhancing bright features.€ Signalout= K • Signalin−γHistogram Levels & EqualizationOne alternative/complementary procedure to manual adjust of brightness/contrast is equalization, which can be applied to the raw image. It stretches out the histogram, with the distinction that it separates the intensities into weighted bins, so that if there are a lot of pixels piled in a few bins, these bins (intensities) will have a larger number of new intensities mapped onto them – i.e., there will be “spaces” between them on the histogram, meaning those intensities will be stretched out. At the same time, bins with not many pixels in them may be squeezed together, as there is less total information relative to the high populated pixels.Russ, 1999, Fig. 4.11, p. 238.Kernels/Rank OperatorsNoisy images sometimes occur for a variety of reasons, some avoidable, some not. Noise refers to some randomness added to pixel intensity values, with noise worse where count rates are low. The simplest procedure to reduce noise is to take the average of the pixel and its surrounding neighbors, and put this new average value in as the new pixel intensity. You can create a matrix with values for the coefficient by which you weigh (multiply) each pixel and adjoining neighbors. For example, one such matrix could be1 1 1 and 1 2 11 1 1 another 2 4 21 1 1 1 2 1These are called kernels, or rank operators. Say there was a ‘noisy’ pixel with a value of 100, when all the adjoining values were 10. The first kernel would return a new value of 20, and the ‘noise’ would be drastically reduced.Neighborhood averagingResults of applying one kernel: a) A noisy original image, b) each 4x4 block of pixels is averaged (less noise, but too coarse), c) each pixel replaced by average of 3x3 neighbor-hood ( pretty nice), d) each pixel replaced by average of 11x11 neighborhood ( less noise, but too big, causing blurring)Russ, 1999, The Image Processing Handbook (3rd edition), Fig 3.3, p. 166Image MathAbove is an