DL is used to extract and learn features and in the next step to predict for other input images

Just to add to it, DL learns it's own feature and not the handcrafted one used on other hand.

extracting features from images is the purpose of deep learning where as image processing is just process the image in a way that you can find some info. As DL is extracting features or information from lower level to higher level & in every layer it can learn a lot of things that's why DL used in most of the cases. .

It can solve segmentation task together with classification. On current moment, DL approaches already have higher accuracy of segmentation compare to classical

Deep Learning is an application for training data, that may be images, text, sound, numerical etc.

Consequently, analyzing images via deep learning application is again image processing, but indeed it is suitable for large databases of images, sound, numerical etc. Do not be confuse, it is the same thing.

Dear Ayesh Buddhima

This paper may help you

Article Advanced Meta-Heuristics, Convolutional Neural Networks, and...

It is a temporary hype. There is a whole universe of things that can't be achieved by tossing a deluge of sample images into a Google or Amazon service. For example, deep learning systems don't execute user-defined algorithms. They don't even count objects, like nodes on a line, they don't build up data structures. They can't tell whether an object is encircled by another. They can't retrieve information from databases.

Regards,

Joachim

As with anything and as I said on previous post every tool has its place in solving a particular problem of science. Deep learning is a buzzword right now and you should treat the high number of publications as noise. I would suggest to focus more on the problem at hand that you are working on and thinking about which tool focuses on solving the problem in the most optimal manner.

Now neural networks are pattern recognizers that if there is enough data can help you by processing all the data. Once you see the results compare them to the results from traditional image processing techniques and let empirical outcomes determine the way you solve your problem.

Additionally, there is a third option where there is room for both techniques to extract the best possible result. data can include artifacts that might contribute to degrading the performance of the NN and traditional image processing is efficient in this manner.

It has to do with the performance. Currently deep learning techniques is the SOTA, and no need of feature engineering. Designing features is a difficult task which the deep learning trying to eliminate. Classical image processing is not well generalizable, well depends on application area as well. Sometimes deep learning can be a overkill. By using simple classical approach you can solve the problem. Again it depends on the problem you are trying to solve. People usually do hybrid approach that have classical as well as modern deep learning techniques.

Image processing techniques only help to modify images applying algorithms, they do not initiate learning outcomes. When visual attributes of processed images are converted to data models called feature vectors and applied to classifiers, then only decision boundaries can be generated which is part of machine learning. For most medical imaging problems, generating feature vectors manually is time consuming and often requires medical expertise. In such cases deep learning neural frameworks have often been found to generate the data models effectively in an automated manner, and hence they are being extensively used in classification problems to reduce software development time cycles.

Deep learning are the state of art methods that can extract feature by themselves, learn the task in end to end fashion where traditional image processing based methods needs explicit feature extraction. However, it all depends on the problem that you are trying to solve.

Feature extraction capability is higher and it undergoes with huge dataset