How can the classification results for the entire image be predicted?

I noticed that you used the OCN anomaly detection network to determine whether a sample belongs to a known or unknown category. However, the input of OCN is fixed at three dimensions. How do you perform inference on the entire image in this case?