As an image goes through more layers in a CNN, what kind of things do the final 'pictures' (feature maps) represent?

As an image progresses through more layers in a Convolutional Neural Network (CNN), the feature maps produced by the final layers represent highly abstract, high-level, and semantically rich information about the original input image. Each layer in a CNN applies filters to its input to detect specific patterns, and the output of these filters forms a feature map, which is a two-dimensional array indicating the presence and strength of a particular feature at different locations. Early layers detect very simple, low-level features such as basic edges (horizontal, vertical, diagonal), corners, or textural patterns. As the information is passed to subsequent intermediate layers, these layers combine the detected low-level features to recognize more complex, mid-level patterns, for instance, a circular shape, a specific curve, or a grid pattern. The final layers then build upon these mid-level representations. The 'pictures' or feature maps at these deepest layers do not represent individual edges or simple shapes. Instead, they represent semantic concepts, which are meaningful components or entire objects relevant to the task the CNN was trained for. For example, if a CNN is trained to classify animals, a feature map in a final layer might strongly activate in regions corresponding to a 'cat's face', a 'dog's paw', or an 'animal's eye'. These final feature maps capture highly distilled representations of the input, indicating the presence and location of the complex object parts or entire objects that define the categories or concepts the network has learned to recognize.