Following

Convolutional neural networks (CNN) are typically the best way to address prediction tasks with images as input. For a regression case, a CNN can be trained to predict the desired scalar output for a given 2D input. 

One thing to keep in mind is that CNNs are typically data-hungry: a large dataset is needed to train the network (the needed data size is dependent on the complexity of the network architecture)

Thanks a lot@ Mohamed Baker Alawieh .

I actually tried CNN, but it seems to possess no convergence, and I tried and failed in finding some analogous examples on Google.

Could you kindly suggest some examples that maps a 2D image to a scalar?

Thank you again.

Hi Yuheng Zheng,

One common regression problem based on images is crowd-estimation which maps an image to a scalar (number of people in image). There are many papers and resources available online for this particular problem.

Here is one implementation example using Keras in python (code not mine):

https://github.com/alphinside/crowd-estimation-cnn-regression

While I did NOT try running the code, I think the network structure shows a good example of a CNN regression. Compared to CNN classification, the main difference is having a single output neuron with, typically, mean squared error loss.

Hope this helps!

Regards,

Mohamed