Following on from my comment above, to get an idea of the broad number of topics that solutions to this question can touch on you just need to consider a simple example where you have images of cars and images of houses and you want to be able to identify existing and new images.
Let’s say images are not the same size. Then you will want to change the image size to some common size. New and existing pixels will have to be predicted from existing pixel values. You may use anything from a simple linear interpolation to a complex nonlinear transformation to determine new pixel values.
To determine if an image is a car, for example, or a house you will need to find certain characteristics of each that are maybe unique (unlikely when you scale the problem up to include more types of images) or that best separate the images. Now you are in the realm of classification. You might start off with a range of typical images, find values for a set of defined characteristics and then use machine learning to train a solution. Again, many choices exist.
When dealing with image the term features is used to describe actual regions on the image. Characteristics better describes image based measurements or statistics
Finding a set of features that best distinguish image types is another interesting problem. Given a super set of features, you may want to reduce the set to a more manageable size (feature selection) and use this reduced set to train and build a classifier. This can take you into areas such as bootstrapping or resampling or even evolutionary algorithms simply to find a best set that is robust from a number of competing sets of features.
Taking a step back, to find any features you will maybe have to investigate image segmentation or feature detection. You will have to detect corresponding features on all images and possible use a feature matching technique to match corresponding features across images. This in turn might act as a starting point for image alignment. Each area mentioned is a full of challenging problems and therefore very rewarding.
In short, your approach might be to
1/ find a set of interesting features or characteristics
2/ use these to align your images
3/ find a subset of interesting features that are robust when used for image classification
4/ Develop a learning model that can classify your images with high sensitivity and specificity.
5/ Apply the above steps to new images to classify as, in this example, a house or a car
Of course, this is but one approach or many possible but is does highlight many of the issues you will face when trying to solve problems such as that described in the question.