An image is a tensor whose axes are the picture.

Demonstration

A model never sees a picture. It sees an array, and the array has a shape:

(H, W, C)      one photograph
(B, H, W, C)   a batch of them
(B, D, H, W, C) a CT volume, with depth

Each entry is a measured intensity at a location in a channel. That is the whole of the representation, and everything called computer vision is arithmetic on it.

Two facts about the shape are worth stating because both cause real bugs.

Channel order is a convention, not a property. PyTorch expects (B, C, H, W); TensorFlow and almost every image library expect (B, H, W, C). Loading with one and computing with the other produces a tensor of the right rank and the wrong meaning. If the height and width happen to match, nothing errors.

Resolution is a choice, and it is made twice. Once by the sensor, and again by whatever resize sits at the head of the pipeline. Downsampling a 4000 × 3000 fundus photograph to 224 × 224 discards roughly 99.6% of the pixels. For classifying a whole eye that may be fine. For finding microaneurysms — lesions that are a few pixels across at full resolution — it is the experiment.

The channel axis is where the modalities differ. Three for RGB. One for a greyscale radiograph. Four when a mask is carried alongside. Many for a multispectral or multi-sequence acquisition, where each channel is a different measurement of the same tissue rather than a different colour of the same light.

Corollary

Resolution and input size are different numbers, and the gap between them is a decision nobody records. Before comparing two vision results, ask what each one resized to. A model reading 512 × 512 and a model reading 224 × 224 are not solving the same problem, whatever the table says.