Connected Morphological Operators Improve Image Classification
Image classification is an active area of research with many applications, including content-based image retrieval, object recognition, and surveillance. Images can be classified based on the presence of a specific object or on certain global properties of object patterns present in the image. Size and shape descriptors are computed from the contour or the area of the objects. Often, these descriptors require considerable computing time. Even worse, many of them are not scale and rotation invariant, which further complicates processing.
We present an efficient method for classifying gray-scale images, based on both size and shape properties of a single object, or on patterns of objects present in the image. Area (defined as the number of pixels) and elongation are used as size and shape descriptors, respectively. The invariance of these elements to translation, scale, and rotation means that all of the desired objects are recognized in one single pass of our algorithm. We compare the classification performance of our method with existing methods.
Pattern spectra Mathematical morphology is a field of research where mathematical concepts are used to describe images and define operators on images.1 Commonly, operators using so-called structuring elements (SEs) are used??small images containing part of the object of interest. The image is then ??probed?? with an SE such that image areas where the SE fits are preserved, while areas where it does not are removed. For each given basic shape, the input image is filtered with SEs of that shape at increasing scales. Such a set of operators, called a granulometry, is used to compute a pattern spectrum??a histogram in which each bin contains the amount of image detail that is associated with the scale and shape of the corresponding SE.2,3 The amount of detail that was removed by each operation at a certain scale is stored in a separate (increasing) bin of a table. The pattern spectrum is then obtained by taking the derivative of that table by size. Finally, the spectra are combined into a single feature vector. The decision-tree classifier C4.54 is then used to evaluate the vectors.
Our approach is based on another technique from mathematical morphology called connected filtering. We demonstrate that its classification performance is equal to or better than the best existing method using SEs. Moreover, the computation time of our method does not depend on the number of scales and shapes used. Our approach has also been compared with several other methods for image classification,5,6 such as a combination of gray-level co-occurrence matrix and Gabor wavelets. For a set of diatom (algae) images, it was found that our method also outperforms these.