What is the minimum dataset required for one class?

I'm doing project on electric circuit diagram recognition. I have segmented the components but after feature extraction the classifier does not classify correctly. I have considered 8 images for each class. Is it enough?

Answer 1

[Ameer Hamza]

Definitely now enough. Not even close. Consider the MNIST dataset, which is used to train a network on handwritten images for recognition of digits, a task much relatively simpler than classifying a circuit diagram. And even it uses 60000 images for training (about 6000 for each digit).

My suggestion is to use some non-image based classification technique. For example, develop a strategy to extract features of a circuit diagram based on its properties, e.g., number of component, type of connections, type of sources, etc. However, this task in itself can be very complicated.

[ABHINAV M S]

I have extracted LBP features along with eccentricity, orientation,extent and Euler number for training image of particular class to svm.

[Ameer Hamza]

Even if you are using SVM, image-based features will still require a large dataset. Even non-image based features will still require quite a large dataset. Also, if you are going to use SVM, then are you sure that these image-based features you mentioned have a strong correlation with the class of circuit diagram?

[ABHINAV M S]

Ok I will through that.Thanks for your suggestion.