I have a problem with Image Classification.
Concretely, I am building an image classifier that could classify 50 classes. In which, there are some classes those are large intra-pattern variance. For example, class "electronics" has images that related to "cooker", "hairdryer", "drill", "sewing", "tv", "washer", etc. These sub-classes are different in pattern. Therefore, when I train, the classifier can not learn well to classify images belonging to "electronics".
Furthermore, I cannot break the big class "electronics" into small sub-classes because I will make the dataset imbalanced due to the number of images belonging to a sub-class is small, comparing to other big classes.
So, anyone can help me to suggest a method to deal with this situation? In my view, a classifier using semantic at the top of the CNN can help me overcome this situation. Because, although sub-classes are different in pattern, they are similar in semantic. However, I have not found out any source related to my idea.
Can you suggest me any solution?
