Chapter 20

Title: Image Mining Extension for RapidMiner (Introductory)


Chapter 20 introduces the RapidMiner IMage Mining (IMMI) Extension and presents some introductory image processing and image mining use cases. Chapter 21 provides more advanced image mining applications.

Given a set of images in a file folder, the image processing task in the first use case in Chapter 20 is to adjust the contrast in all images in the given folder and to store the transformed images in another folder. The IMMI extension provides RapidMiner operators for reading and writing images, which can be used within a RapidMiner loop iterating over all files in the given directory, adjusting the contrast of each of these images, for example, using a histogram equalization method. Then the chapter describes image conversions between color and gray-scale images and different feature extraction methods, which convert image data in unstructured form into a tabular form. Feature extraction algorithms for images can be divided into three basic categories: local-level, segment-level, and global-level feature extraction. The term local-level denotes that information is mined from given points (locations) in the image. Local-level feature extraction is suitable for segmentation, object detection or area detection. From each point in the image, it is possible to extract information like pixel gray value, minimal or maximal gray value in a specified radius, value after applying kernel function (blurring, edge enhancements). Examples of utilization of such data are the trainable segmentation of an image, point of interest detection, and object detection. The term segment-level denotes feature extraction from segments. Many different segmentation algorithms exist, such as k-means, watershed, or statistical region merging. Segment level feature extraction algorithms extract information from the whole segments. Examples of such features are mean, median, lowest and highest gray value, circularity, and eccentricity. In contrast to local-level features, it does not take into consideration only a single point and its neighborhood, however, it considers the whole segment and its properties like shape, size, and roundness. With the use of knowledge about the size or shape of target objects, it is for example possible to select or remove objects according to their size or shape. The global-level denotes feature extraction from the whole image, for example, mean color, dominant color, maximal gray value, minimal gray value, variance of pixels, number of edges etc. Unlike the local or segment level, the global level segmentation is not suitable for points or areas identification or segmentation. Rather, it is suitable for classification of images and determining properties of the image as a whole. Chapter 20 provides examples demonstrating the use of local-level, segment-level, and global-level feature extraction. Local-level feature extraction is used for trainable image segmentation with radial-basis function (RBF) Support Vector Machines (SVM). Segment-level feature extraction and trainable segment selection reveal interesting segment properties like size and shape for image analysis. With the help of global-level feature extraction, images are classified into pre-defined classes. In the presented use case, two classes of images are distinguished automatically: images containing birds and images containing sunsets. To achieve this, global features like dominant color, minimal intensity, maximal intensity, percent of edges, etc. are extracted and based on those, an image classifier is trained.

Table of Contents

20.1 Introduction
20.2 Image Reading/Writing
20.3 Conversion between Colour and Grayscale Images
20.4 Feature Extraction
20.4.1 Local Level Feature Extraction
20.4.2 Segment-Level Feature Extraction
20.4.3 Global-Level Feature Extraction
20.5 Summary
20.5 Exercises
20.5 Glossary
20.5 Bibliography

Data & Processes: Click here to download