Deep Learning for Maritime environment

Hi everyone,

I want to realize a ship detector based on neural network. It is the first time that i try to use neural network for object detection therefore my question may be trivial. My idea is to use transfer learning on Alexnet and than to train the detector with this network.

DATASET:

• How many images a need in order to train perform the transfer learning on Alexnet?
• Should the detector be able to recognize ships seen from every point of view (if images with multiple point of view are included in the data-set) or it is necessary to use multiple detector to identify different ship observed from different orientations (HOG detector style)?
• Should the images contain only the object to be detected (ships in this case) or the entire maritime scene with the ship highlighted by a boundig-box?
• I found online two or three data set that correspond at the descriprion of the first case above, but anyone for the second case. Any link suggestion for the second case? (that is the one that I think is correct)

OPTIONS:

I found multiple type of options navigating in the web, the most common used are:

opts = trainingOptions('sgdm', ...
    'Momentum', 0.9, ...
    'InitialLearnRate', 0.001, ...
    'LearnRateSchedule', 'piecewise', ...
    'LearnRateDropFactor', 0.1, ...
    'LearnRateDropPeriod', 8, ...
    'L2Regularization', 0.001, ...
    'MaxEpochs', 100, ...
    'MiniBatchSize', 256, ...
    'Verbose', true);

How can i relate the MiniBatchSize to the amount of images that I have as dataset?

Thank you!

Answer 1

[Gabija Marsalkaite]

The number of images is a difficult question. The smallest dataset I trained a network on was around 1000 images but that depends on how easy to detect the object and if the network was pretrained on similar images. You don't need separate trained networks for the detection of an object if you have sufficient number of different side examples in training set. My recommendation for minibatch is the biggest what you can fit into your hardware memory (GPU memory or RAM) before getting out of memory errors. I've heard recommendations on choosing power of 2 if you are using a GPU but did not test that myself.

For ships highlighted by a bounding-box I would suggest looking on ImageNet:

http://image-net.org/download-bboxes

Alexnet is not first recommendation because of its low accuracy, you can see comparison in link below:

https://www.mathworks.com/help/deeplearning/ug/pretrained-convolutional-neural-networks.html

If you don't want to write everything from scratch, these examples may be useful:

https://www.mathworks.com/help/vision/examples/object-detection-using-deep-learning.html

https://www.mathworks.com/help/vision/examples/object-detection-using-faster-r-cnn-deep-learning.html

https://www.mathworks.com/help/vision/ug/train-an-object-detector-using-you-only-look-once.html

[VaL V6]

Thank you, your answer will be very helpful. Since Alexnet has low accuracy, looking at the link you propsed, I'll try inceptionv3 that it seems to represent a good trade off between accuracy and prediction time.

Then, if I want to try with a scratch network, do you have any suggestion about the architecture?