What you are required to do is to run one of these types of deep learning algorithms on a data set you created (transfer learning).
There are quite a few websites that you can use to train and test your model. For example, YoloV5 for object detection, Unet semantic segmentation, or other existing deep learning applications such as CLIP, inpainting, Siamese networks etc.
For object detection you take a trained model and train it to the classes you want. You can then test the algorithm on some test images. The algorithm detects the objects in the images returning a set of bounding boxes their classes and the probability that the object belongs to the class.
So choose a few (2-3) classes you want to detect. Take a few dozen images in which the objects appear. Divide them into train and test. Label the images using a labeling tool and then train the model for your classes. Once the algorithm has converged generate a program to run the algorithm on the test images.
For semantic segmentation take several images and annotate them where each pixel belongs to one of 2-3 classes.
You can also use other types of algorithms but you will need to perform some type of transfer learning
Maybe think about a post processing step you want to run on the output to yield a higher level of understanding.
At the end write a report on your project and show in it some results (good and bad).
If for some reason you need to do a similar project for your final project or your thesis you can use the same data for your EX.
Good Luck.