Project Title- Digital Image Processing: Enhancement, Segmentation and Interpretation
CS 586- Spring 2008
Introduction
Project Image
Enhancement Techniques
Segmentation
Interpretation
Conclusion
References
Introduction:
Digital image processing is the process of subjecting a numerical representation of an object to a series of operations in order to obtain a desired result. The processing changes the form of the image to make them more desirable or attractive, or to accomplish some other defined goals. Castleman (1996) defined the process as a sampled, quantized function of two dimensions that has been generated by optical means, sampled in an equally spaced rectangular grid pattern, and quantized in equal intervals of amplitude. The term digital image processing is often used to cover both processing and analysis. The procedures are normally categorized into enhancement techniques, segmentation and interpretation. Gonzalez and Woods (2006) describe digital image processing as the processes of acquiring an image of an area, preprocessing that image, extracting (segmenting) the individual pixels, describing the pixels and recognizing the elements in the image. Image enhancement is the process of making an image more interpretable for a particular application (Faust, 1989). Usually the goal of image enhancement is to improve the visual interpretability of an image by increasing the apparent distinction between the features in the image. Segmentation consists of subdividing an image into its constituent parts as well as extracting them and tends to be a key issue in pattern recognition and image understanding. Image interpretation involves assigning meaning and classes to segmented units of pixels in the image. Application of digital image processing ranges from industrial quality control, medicine, robot navigation, geophysical exploration, remote sensing, and military applications.
The study will apply the digital image processing techniques (ie. enhancement, segmentation, interpretation) to a remotely sensed data.
Project Image:
The image understudy is a Land Satellite Thematic Mapper 5(Landsat TM 5) image acquired in 16th August, 2006. The image is composed of 7 different channels (bands) cutting across the visible and the infrared spectrum of the electromagnetic spectrum (EMS). The data span across different wavelength and frequency. The image covers an area of Southern Mississippi. For this application a subset was acquired to speed up processing and reduce computational time. The subset covers an area demarcated for a new dam and lake construction for recreational purposes
.
 Fig 1a: Original gray level image data |
 Fig 1b: Original multiple bands image |
The enhancement techniques explored in the image consist of contrast manipulation, spatial feature manipulation, and multi-image manipulation (Lillesand et al, 2008; Gonzalez and Woods 2006; Aronoff, 2005; Davies, 2005; Jensen, 2000).
 Figure 2a: Original image |
 Figure 2b: Linear contrast stretch |
 Figure 2c: Piecewise contrast stretch |
 Figure 2d: Histogram equalized stretch |
 Figure 3a: 3x3 low pass filter |
 Figure 3b: 3x3 high pass filter |
 Figure 3c: convolution 5x5 low pass kernel |
 Figure 3d: convolution 5x5 high pass kernel |
In figure
2b the linear contrast stretch is applied to the image to improve the visible
contrast of features. The original image data may fall within a narrow range of
pixel values. As a result, it may hinder visual display of distinguishable
features. When the linear contrast stretch is applied the total range of values
are spread across the brightness values 0 to 255. The algorithm can be specified
as: 
where
DN is equal to digital number assigned to pixel in output image and DN is the
original digital number of pixel in input image. Features can be discernible as
result of the stretch. To enhance shadow and areas of low contrast the piecewise
contrast stretch is applied to the original image (Fig 2c). At times the
histogram of the image data may not be Gaussian and piecewise contrast
stretching can be very useful. Figure 2d is a histogram equalized stretch of the
original image. The histogram equalization is a nonlinear stretch that
redistributes pixel values so that there is approximately the same number of
pixels with each value within a range. Contrast is increased at the peaks of the
histogram and lessened at the tails.
 Figure 4a: Laplacian edge detect |
 Figure 4b: 7x7 edge detect |
 Fig 4c: RGB-original image |
 Fig 4d: RGB contrast stretch |
 Fig 4e: Histogram Equalized stretch |
 Fig 4f: Sobel nondirectional edge detect |
 Fig 4g: Vegetation Index-NDVI |
 Fig 4h: PCA 1 output |
A variety of segmentation algorithms have been developed in the last few decades that can be readily applied to all images. Some of these algorithms may not be suitable for some particular situations, especially in the case of satellite imagery, which, often, contain different textured regions or varying background, and are often subjected to illumination changes or environmental effects. The techniques explored (threshold, edge, region, watershed) still have disadvantages and add to the fact that there are no perfect segmentation algorithms. In satellite image segmentation, the region-based techniques tend to be very popular and very useful (Pohl & Genderen, 1998; Gonzalez and Woods, 2006; Lillesand et al., 2008).
 Figure 7a. Original Image |
 Fig 7b. Segmentation results with growing regions |
Figure 7b indicates the segmentation results with the growing region process. The regions take into account global information as well as some local information relative to the pixel. The algorithms take one or more pixels, called seeds, and grow the regions around them based upon a certain homogeneity criteria. If the adjoining pixels are similar to the seed, they are merged with them within a single region. The process continues until all the pixels in the image are assigned to one or more regions. When there is not any neighboring pixel which is similar to the region, the segmentation of the region is completed. Figure 8b shows an application of texture analysis to segment and mask out distinctive regions.
|
Figure 8a. Original image |
 Figure 8b. Segmentation based on texture analysis |
The result of the texture based segmentation produced only a few regions. The regions generated are very interesting for picking unique regions. Figures 9b and 10b depict a color coded (RGB) segmentation based on thresholding followed by regionalization .
 Figure 9a. Original image |
 Figure 9b. Color coded segmentation |
 Figure 10 a. Original image |
 Figure 10b. Segmentation result after thresholding and region growing process |
back to the top
Development of a rigorous and complete set of class definitions is critical to the success of an interpretation. In a satellite image the information content represents cover features. It has become a de facto standard for satellite image interpretation to label classes at the first and second levels of Anderson’s classification system (Anderson et al., 1976). Usually the classes are referred to as information classes. The process of assigning useful descriptions to segmented regions rest on the image analyst and can be time consuming. Interpreting satellite image also requires field validation of classes identified. Figures 11a and 11b denote the final labeling in the interpretation process based on the segmented regions (figure 10b).
 Figure 11a: Segmented results (color-coded) |
 Figure 11b: Interpreted results (legend) |
back to the topIn all, four major classes were identified. The classes were labeled as mixed forest land, cropland and regrowth, nonforested wetland, and waterbody. Information such as texture, shape, digital number values (DN), spectral characteristics, and background knowledge were used to interpret and label the regions. Waterbody class was easily identified and labeled because of the polygon shape and the texture of the feature class. In addition, the digital number value of the pixel is unique to other landscape features. Texture, DN values and spectral characteristics played major role in labeling the mixed forest land, the cropland and regrowth, and the nonforested wetland classes. Shape played little or no role in these classes because of boundary problems. Knowledge of the image area and existing information also played a major role in region identification and labeling as classes.
Digital image processing methods have become essential for both visual and digital analysis of images including remotely sensed data. Image enhancement techniques are used to improve the interpretability of the image product, segmentation techniques are used to identify and delineate regions and interpretation is done to achieve final labeling of classes in the image. The results of the project show the various stages of the image processing process. Even though the concept of the process is the same for diverse discipline of images, the applicability and the results always depends on the background, experience and expert knowledge of the image understudy. In other words, there is no one size fits all in digital image processing.