I impemented entropy measures (Renyi, Shannon, Havrda & Charvat and Kapur) on infected red blood cell images.
How can we conclude which image is infected not infected based on entropy measure values? or How to predict (conclude) based on entropy measure values (higher or lesser)?
Generally speaking you have two class A (infected) and B (not infected). All entropy measures will indicate more ordered states with lower values and disordered with higher. From data (images) described with appropriate parameters (features) at disposal classification entropy threshold value can be detected which is on the other hand is not an easy task - some good heuristic or expert knowledge should be used.
Dear Malarvel , my view points here given as follows :
1. First of all Entropy means : Amount of redundant information (amount of information contains in image)
2.I hope your application is classification of infected and not infected , so you may try any classification tool like neural network or any modified classifiers
3 In case of Classification , role of Entropy is as the one training parameter
4. But If you want to concentrate only in Entropy then you suppose to do training with many sample images [ both infected and not infected]
5. If you use more parameters accuracy will increase
I hope it will helps you , for future any queries [email protected]
I will try to help, if it is possible, with some additional explanations. First, we all need to agree on same that this is so called supervised case - if it so, than more images correctly classified as infected and not infected is undoubtedly a preferable prerequisite. From here we can define two class problem. Classification can be done by incorporating different parameters derived from images. Because we do not know, in advance, deterministic creation mechanism of the data described with defined parameters they are observed through random variables. In literature various definitions of the entropy as the information content measure can be found. But, for classification possible usage of entropy can be folowing. Estimate entropy from your data, separately for class A and class B (this is possible in the supervised problems) --> with defined parameters (eg. 10). Here is the moment where discriminative power of the entropy could be misleading if the obtined values for class A and class B are pretty close to each other. This can have lots of causes - like noisy data, high redundancy, engaged parameters with low discriminative power, etc. On the other hand for cases in wich entropy values for class A and B is obviously different next step is to use them on unseen data (images). Add new image of unknown class to your data set in a way to join it to the class A and estimate entropy for that case. After that repeat all the same for the class B. Now you have delta of the entropy for two cases. If the delta of one case is much smaller than in other case that means your new data / image can be classified in a particular class with smaller delta. Problems can occur when delta of the entropy is similar for both classes - threshold for delta values are need to be defined by good heuristic or expert knowledge. Also, when great overlapping is detected in the data set probably described procedure would not be good choice in the classification problems. Hope this will be usefull in some way.
Regards
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing