For an instance we capture images with random sizes, orientation, filters etc., How will you decide whether the captured images can be used for deep learning applications ? Is there any standard to be followed ? Or any procedure to be followed while capturing images ?
Not only for deep learning, for any image processing / pattern recognition problems images need to be standardized with respect to type, size, resolution, orientation etc. as the features computed are usually dependent on these, unless they are specifically made to be invariant to these parameters by some normalization process
It is good the images to be the same resolution and format (for say .jpeg), and represent mainly the object you want to train your ML model to recognize, it is same for every class of objects if you are doing classification.
About classifying the different objects, it's enough to put the images of every object in separate folders and name them - 1,2,3...n, if you are taking and photos of each detail of an object, put them in subfolders. The rest will be done by the software, for example Tensorflow has dataset generator function, which identifies the classes and subclasses automatically based on the folders and subfolders.
About taking the photos with different orientation and filter, it's excellent cause that way your ML model will be able to learn to recognize the object under any condition.
And about the different resolution of the images, this is useful for validation or test set, cause that way you will be able to validate you model on never before seen images, making sure that you are not overfitting or underfitting it.
Also don't worry if your images are not the same resolution or if you are using the same filter, for example Tensorflow has build in image augmentation function, which will randomly change the images fed into the model.
- Badges
- Science topic