what are the ideal requiremnets for data which should be kept in mind when data is collected/ extracted for Image classification. What are different techniques in machine learning or deep learning that could be applied.
what should be pixel size ? resolution ? how many number of images does each algorithm need ?
I think, the important things is
1. Amount number of data
2. Balance number of data for each class
3. Variation of data
The exact number of data required is unknow but usually more is better. Some problem can converge with 1000-2000 data with fine tuning strategy. Maybe you can also use fine tuning strategy for your problem.
It really depends what you want to do. Dataset may vary according to your requirements. If you want to implement face or facial expression recognition, you can use JAFFE, CK, CK+, RaFD, MMI, ORL etc.
1. Some paper recommends 128*128, some recommend 64*64. It depends on your work.
2. It should be better if you have a large amount of data. If you use a greater amount of data for training, your recognition accuracy should increase.
You should follow the following papers:
https://www.researchgate.net/publication/316906235_Weighted_score_level_fusion_of_iris_and_face_to_identify_an_individual
https://www.researchgate.net/publication/317011770_Recognition_of_an_Individual_using_the_Unique_Features_of_Human_face
https://www.researchgate.net/publication/315511448_Human_iris_as_a_biometric_for_identity_verification
https://www.researchgate.net/publication/305297599_PCA_and_back-propagation_neural_network_based_face_recognition_system
https://www.researchgate.net/publication/308824012_Face_recognition_using_Principle_Component_Analysis_and_Linear_Discriminant_Analysis
Empirical suggestions
1. Amount. For machinie learning: it should be at least 10 times the number of features per class . For Deep Learning it should be at least 100 times the number of features;
2. The sample quantities should be balanced among classes;
3. The samples should represent the real situations where you are going to apply your model. For instance, if you are training a face classifier for using in situations where faces are smaller than 30x30 pixels you should have low quality and low resolution face images in the training stage;
4. Samples should have the maximum variety possible. For instance, in face classification, your dataset must have faces of people in different ages, ethnicities, genders, illumination conditions, orientation in-plane and out-of-plane, etc.
Eanes Torres Pereira could you please share a research paper or citation that mentions about the 10 times number of features of class and 100 times the number of features of class, as it would be helpful in our research.
These rules are mostly empirical. But I think these papers may have some useful information: Article Sample Size Planning for Classification Models
Article Predicting sample size required for classification performance
https://sci2s.ugr.es/keel/pdf/specific/articulo/raudys91.pdf
Article Dimensionality and Sample Size Considerations in Pattern Rec...
Article Sample Size Planning for Classification Models
- Badges
- Science topic
- Similar topics
- Computer Science
- Images