I need to increase the input resolution of a "simple" LeNet CNN, from 32x32 to 640x480. The net must perform object classification, therefore I need probability values as output (scalar quantity). I could perform a big spatial subsampling (increasing the stride factor in the pooling layers), but I don't know how much the accuracy of the classification is affected by this choice. Is it better to add layers, making a deeper CNN?
LeNet was trained on the MNIST dataset, right? This is a dataset of handwritten digits, so I'm doubtful that this could be used straight from the box with object recognition. Or perhaps you mean just using the same kind of architecture?
To give some background/context, what you want can be tested easily enough with libraries like Keras (https://github.com/fchollet/keras/). The following is a good starting place (https://github.com/fchollet/keras/blob/master/examples/cifar10_cnn.py). You would need to adjust the architecture specification. In the cifar10 link, the input images are 32x32, and this is set in the code with the variables "img_rows" and "img_cols". For your case you would want to set these to 640 and 480. To perform convolutions, you need to inspect the "Convolution2D" instantiations. On line 45, for example, 32 convolution filters of size 3x3 are defined. You might want to increase the filter size in your dataset. Perhaps the new filter sizes would need to be 32x32 or 64x64? Finally, max pooling is done with "MaxPooling2D" instantiations (e.g. at line 50).
I'm afraid there probably aren't any general answers to your questions, and the community would benefit for an extensive empirical investigation in this. This is a general problem with neural networks. It's generally not possible to predict how well a given architecture will perform on a given dataset without an empirical investigation. But to address your specific questions, the point of max pooling is to reduce the spatial dependency of the network. So, if you do this aggressively, I suspect that a lot of information would be lost, so I would be wary of doing this too much. On the topic of depth, recent networks have no fear of depth. For example, Microsoft research won the image net challenge last year with a CNN with 152 layers (http://blogs.microsoft.com/next/2015/12/10/microsoft-researchers-win-imagenet-computer-vision-challenge/)! So perhaps this is an option you can investigate. Read recent works to make sure you do this appropriately.
Ok, thank you for the reply. I want only use the LeNet architecture (or something of similar complexity), but I will train it on a specific dataset for object recognition.
I did a quick search, and couldn't find any recent explicit studies regarding the effect of max pooling and network depth on object recognition. I recommend you do an extensive empirical analysis and publish in a good conference. It would be a valuable contribution!
- Badges
- Science topic