Performing accurate measurements on objects in images is fundamental to solving many image analysis problems. Our group has considerable experience in finding novel measures which extract the maximum information from the available data. These measurements may be simple, such as the number, size, or colour of objects, or more complicated, such as the shape, connectivity, or appearance (texture) of objects (there are over 100 different measurements describing image texture alone). Additional measurements might describe the spatial arrangement or distribution of objects in a scene, or the statistical distribution of properties across many objects. Sometimes it is well established beforehand which features of the image need to be measured, at other times suitable measurements are “discovered” from a large number of pre-computed possibilities.
For example, in the Melanoma Diagnosis project which sought to classify skin lesions, over 600 features were considered for inclusion in the final algorithm. These were later culled to the 80 best based on performance, and later to less than 12 for the final model.
In other cases, a specific operator can be constructed to measure some particular property of the data – for example the dark spots (melanocytes) on the boundary of a melanoma can be detected and measured by a novel boundary-restricted top-hat operator.
There are many classes of features and each has various techniques for measurement. In addition higher order features are formed by combinations or distributions of the simpler measurements. For example:
- object size (area, volume, perimeter, surface) – obtained by counting pixels
- object shape – obtained by characterising the border – Fourier descriptors, invariant moments, shape measures, skeletons, edge abruptness
- object colour – description in colour-space, integrated optical density, absolute and relative colours
- object appearance/texture – colour variation in pixel neighbourhoods – co-occurrence matrices, run lengths, fractal measures, statistical geometric features
- Parameters from fitted statistical models – used within objects for texture, eg. Markov Random fields – or to describe placement of objects within a scene (eg Poisson models)
- Distributional parameters – moments: mean, variance, skewness, kurtosis, median, inter-quartile range – used to describe statistical distributions of the more fundamental features, for example within a scene.
We have libraries of software available to implement all these measurements and many more on image data.
Feature extraction was an important part of the following projects:
- Melanoma Diagnosis
- Food Quality
|Image Capture||Feature Extraction|