Hi

Feature extraction could be defined in the territory of the pattern recognition, where you are going to detect a pattern from a massive volume of the data, probably overwhelmed with noise. The new pattern could be actually perceived and constructed with some properties. One might take them as the properties, i.e. features, to be compared with the previously learned patterns, e.g. if the goal is detection of a lesion pattern in a bio-medical stuff, the reasonable feature could be the transparency of the acquired pixels. The analysis to recognize the pattern is totally applying on the raw input to gain the features. It is suffice to say that, the features will be structured in some arrays for analytic computations.

You can get some more coherent ideas with reading the technical textbooks in Image Processing and Pattern recognition.

Best

Hi

Feature extraction could be defined in the territory of the pattern recognition, where you are going to detect a pattern from a massive volume of the data, probably overwhelmed with noise. The new pattern could be actually perceived and constructed with some properties. One might take them as the properties, i.e. features, to be compared with the previously learned patterns, e.g. if the goal is detection of a lesion pattern in a bio-medical stuff, the reasonable feature could be the transparency of the acquired pixels. The analysis to recognize the pattern is totally applying on the raw input to gain the features. It is suffice to say that, the features will be structured in some arrays for analytic computations.

You can get some more coherent ideas with reading the technical textbooks in Image Processing and Pattern recognition.

Best

Globally speaking, information extraction can be conducted based on :

1- "Raw" pixel values

2- Different measures derived from these "raw" pixel values. These measures can reflect geometric information such us shape, spatial information such us texture or topological information, ...etc. These measures are usually called FEATURES.

The previous answers pretty much define features. In addition to the answers, I can give a number of simple to complex examples of features for clarity:

1) Edges from any image represent features. Thus, any edge detector is a feature detector. Example: Canny, Sobel, Prewitt, Laplacian-of-Gaussian etc.

2) Corner detectors also represent feature detectors. Example: Harris, KLT, FAST, SUSAN etc.

3) Complex feature detectors that detect UNIQUE features from any image. UNIQUE because these features are "invariant" to scaling, rotation, translation, illumination etc. Example: SIFT, SURF, ORB, BRISK, DAISY etc.

Unique features represent some specific details PARTICULAR to a specific image. Thus, using features instead of raw image pixel values, would

1) reduce data size (speeding up the process for classification)

2) improve performance (due to use of unique features and redundant data)

I hope this clears up the use of features, a bit.

Features are what you use to convey the high level knowledge the expert has on the problem at hand to the classifier. For an example, if the problem is detecting golf balls on a field (a silly example).. A feature for an example could be the color of each segmented blob. Of course real problems involve in complex problems with hundreds of features... But I just wanted you to get the idea.

Feature is a smallest unit which matches maximally with the similar type and minimally with different type of region. Now classification will depend upon identification of features and segmentation of image based on regions of similar type. Watershed algorithm is popular for image segmentation. Now what exactly one wants to classify in image processing. If you want to classify different images like all the images where vehicle is featured on one side or all the images with flowers, trees, ect. Will be classified after segmentation and recognition with desired objects.

Extracting features from an image is done often in order to classify tissue or objects into different classes. You might for example extract/measure the gray scale median from an area of the brain (healthy) in order to compare it with some other area of the brain (diseased) and establish a relatioship or being able to classify the two areas into healthy and diseased.

Please look into it

www.lifesciencesite.com/lsj/life1004s/091_17045life1004s_593_598.pdf

Article Content Based Image Retrieval by Shape, Color and Relevance Feedback