Deep learning for image segmentation: veritable or overhyped?

Abstract: Deep learning has achieved great success as a powerful classification tool and also made great progress in sematic segmentation. As a result, many researchers also believe that deep learning is the most powerful tool for pixel level image segmentation. Could deep learning achieve the same pixel level accuracy as traditional image segmentation techniques by mapping the features of the object into a non-linear function? This paper gives a short survey of the accuracies achieved by deep learning so far in image classification and image segmentation. Compared to the high accuracies achieved by deep learning in classifying limited categories in international vision challenges, the image segmentation accuracies achieved by deep learning in the same challenges are only about eighty percent. On the contrary, the image segmentation accuracies achieved in international biomedical challenges are close to ninty five percent. Why the difference is so big? Since the accuracies of the competitors methods are only evaluated based on their submitted results instead of reproducing the results by submitting the source codes or the software, are the achieved accuracies verifiable or overhyped? We are going to find it out by analyzing the working principle of deep learning. Finally, we compared the accuracies of state of the art deep learning methods with a threshold selection method quantitatively. Experimental results showed that the threshold selection method could achieve significantly higher accuracy than deep learning methods in image segmentation.