python 2.7 - vehicle type identification with neural network -

i given project on vehicle type identification neural network , how came know awesomeness of neural technology.

i beginner field, have sufficient materials learn it. want know places start project specifically, biggest problem don't have time. appreciate help. importantly, want learn how match patterns images (in case, vehicles).

i'd know if python language start in, i'm comfortable it.

i having images of cars input , need classify cars there model number.

eg: audi a4,audi a6,audi a8,etc 

you didn't whether can use existing framework or need implement solution scratch, either way python excellent language coding neural networks.

if can use framework, check out theano, written in python , complete neural network framework available in language:

http://www.deeplearning.net/software/theano/

if need write implementation scratch, @ book 'machine learning, algorithmic perspective' stephen marsland. contains example python code implementing basic multilayered neural network.

as how proceed, you'll want convert images 1-d input vectors. don't worry losing 2-d information, network learn 'receptive fields' on own extract 2-d features. normalize pixel intensities -1 1 range (or better yet, 0 mean standard deviation of 1). if images centered , normalized same size simple feed-forward network should sufficient. if cars vary wildly in angle or distance camera, may need use convolutional neural network, that's more complex implement (there examples in theano documentation). basic feed-forward network try using 2 hidden layers , anywhere 0.5 1.5 x number of pixels in each layer.

break dataset separate training, validation, , testing sets (perhaps 0.6, 0.2, 0.2 ratio respectively) , make sure each image appears in 1 set. train on training set, , don't use regularization until you're getting close 100% of training instances correct. can use validation set monitor progress on instances you're not training on. performance should worse on validation set training set. stop training when performance on validation set stops improving. once you've accomplished can try different regularization constants , choose 1 results in best validation set performance. test set tell how final result performing (but don't change based on test set results, or risk overfitting too!).

if car images complex , varied , cannot basic feed-forward net perform well, might consider using 'deep learning'. is, add more layers , pre-train them using unsupervised training. there's detailed tutorial on how here (though code examples in matlab/octave):

http://ufldl.stanford.edu/wiki/index.php/ufldl_tutorial

again, adds lot of complexity. try basic feed-forward nn first.