I have the task to predict a numerical value for an image. One way would be to extract some images features by hand (e.g. color histogram) and use these features with a normal regression method. But I am not an expert in computer vision, so I have no idea, what kind of features to use.
However, I was reading about the application of deep learning for classification tasks. I was now wondering if I could use such an network for an regression task as well.
Hi Klausen,
the difference between regression and classification is only in the objective function. You can take any toolkit for classification and just change the objective function.
For example cuda-convnet has Sum-of-squares cost layer which could be suitable for you depending on your regression problem (https://code.google.com/p/cuda-convnet/wiki/LayerParams#Logistic_regression_cost_layer).
You could look at Theano. It is fast, flexible and it has tutorials http://deeplearning.net/tutorial/.
The question is if deep convolutional networks are suitable for your regression task, if you have large training set, ...
Regards.
Hi Klausen,
the difference between regression and classification is only in the objective function. You can take any toolkit for classification and just change the objective function.
For example cuda-convnet has Sum-of-squares cost layer which could be suitable for you depending on your regression problem (https://code.google.com/p/cuda-convnet/wiki/LayerParams#Logistic_regression_cost_layer).
You could look at Theano. It is fast, flexible and it has tutorials http://deeplearning.net/tutorial/.
The question is if deep convolutional networks are suitable for your regression task, if you have large training set, ...
Regards.
Hi,
In order to solve your problem you could have convolution layers with pooling for your feature detection then on top of that a MLP and a logistic regression layer on top of that.
you can take a look at this http://deeplearning.net/tutorial/mlp.html
Regards,
This seems to be an old question, however, I am adding this for future users.
For image classification/detection, Convolutional Neural Networks (CNNs/ConvNets) are strongly recommended. Regression task (for example, finding bounding boxes, or finding location of eyes, etc.), only changes the final output layer while retaining previous ConvNet layers. This allows one to use pre-trained network models to begin with.
Caffe is one of the frameworks one can use. Here is a good tutorial to start with: http://tutorial.caffe.berkeleyvision.org/
For regression, Caffe supports 'SmoothL1Loss'. You can also use the older EuclidianLoss layer for the regression task.
For an actual network that uses regression, look at Faster-RCNN: https://github.com/rbgirshick/py-faster-rcnn
This is a detection network and uses uses SmoothL1Loss for finding bounding boxes as well as region proposals.
Hope it helps!
- Badges
- Science topic