In my works in image processing, i found a variety of images like: x-ray medical images, MRI medical images, satellite images, ...etc. But i could not understand how these images are classified into these different types, i.e.; what are the main factors leading to such classifications? Then, what are the well known types or classes adopted in this field?
Thus, for example, how can we classify the following test-images often used in image processing experiments : lena, barbara, peppers, cameraman, house, mandril, lake, livingroom, ... ?
and thank you in advance.
Actually there exists a lot of different classifications of images, for example, according to wavelength (frequency) range in which they were obtained (radar, optical, hyperspectral, etc.); according to noise type present in such images (medical - speckle, radar - multiplicative or multiplicative+additive, optical - additive or additive+quasipoisson, hyperspectral - additive+quasipoisson, etc.); according to such called image complexity determined by persentage of present in it homogeneous (where intensity level almost does not change) and heterogeneous (containing features, edges, texture, etc.) regions, in such aspect images are usually divided into low-textured (many homogeneous regions), medium-textured and highly-textured (containing few homogeneous regions and great amount of texture and features).
What about standard test images like lena, barbara, baboon, peppers, these are optical noise-free images. In our research group we usually clasify these images according to their complexity: lena and peppers are considered as low-textured, barbara as medium textured, baboon and goldhill - as highly textured ones.
Image classification depends of the task you're solving. It is possible to classify the same images according to number of colors (clusters) present, according to depicted object, according to composition and so on... lots of variants.
If you want to learn about how to design algorithms that would classify images into categories or an image database of already classified images to served as ground truth to train your own algorithms, I suggest you look at the PASCAL VOC (Visual Object Classes) challenges. This will give you a very good idea of what are currently the best approaches.
http://host.robots.ox.ac.uk/pascal/VOC/
http://rodrigob.github.io/are_we_there_yet/build/classification_datasets_results.html
Actually there exists a lot of different classifications of images, for example, according to wavelength (frequency) range in which they were obtained (radar, optical, hyperspectral, etc.); according to noise type present in such images (medical - speckle, radar - multiplicative or multiplicative+additive, optical - additive or additive+quasipoisson, hyperspectral - additive+quasipoisson, etc.); according to such called image complexity determined by persentage of present in it homogeneous (where intensity level almost does not change) and heterogeneous (containing features, edges, texture, etc.) regions, in such aspect images are usually divided into low-textured (many homogeneous regions), medium-textured and highly-textured (containing few homogeneous regions and great amount of texture and features).
What about standard test images like lena, barbara, baboon, peppers, these are optical noise-free images. In our research group we usually clasify these images according to their complexity: lena and peppers are considered as low-textured, barbara as medium textured, baboon and goldhill - as highly textured ones.
Image classification depends of the task you're solving. It is possible to classify the same images according to number of colors (clusters) present, according to depicted object, according to composition and so on... lots of variants.
"Classification" is a too great word: these images are members of "standard test sets" that have been composed for algorithm evaluation. If you "search" a bit around, you may be able to find the complete test sets to which they belong. There you may be able to find for which purpose the test sets were originally generated.
As they are easily available, they spread and sometimes "abused" for tasks they were originally not target at.
Hope this helps a bit.
simple classifiers are metric distance. Efficace classifier is SVM. There is NN,Bayes,Random Forest...
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing