kristoffzoghbi / CnnForAndroid

The Convolutional Neural Network(CNN) for Android

##CnnForAndroid:A Classification Project using Convolutional Neural Network(CNN) in Android platform。It also support Caffe Model

CnnForAndroid is a android platform's implementation of deep learning using Tiny-cnn structure and provide two Recognition sample:one is gender Recognition for caffe net ; two is Car logo recognition for tiny-cnn net.

#Todo List

• add opencl support.
• change to the tiny-dnn new version
• Optimise the code and improve the speed.

#Dependencies

Opencv(for Android platform Opencv-2.4.9)

Tiny-cnn(old Version)

protobuf

#Supporting Caffe model

tiny-cnn provide the caffe-convertor.cpp to support the caffe model.The project also support the caffe model from compiling the caffe_convertor and protobuf.

#For Gender Recogniton

this project also provide a sample for caffe model to distingguish man from woman also called gender recognition.

1.Where from training data?

MORPH Album 2.

the test accuracy is 90.01% in my caffe's net.

2.the net of caffe ?

3.How to train yourself caffe's model?

(1)Please using caffe and train your model.     
(2)then replace  /assets/tinyfile//.caffemodel and /assets/tinyfile/*.protobuf file/.
(3)Finish  change those filenames in tinyCnn.java file.

#How to install it?

no need install.

Just download or git clone it and then open it by eclipse with opencv for java lib ,  use NDK to build it.
Surely your libs docu have the opencv_java.so file or you must add this file(from Opencv-android docu).

APK Download

#Compile it

needful tools： NDK + eclipse + adt.

or android studio.

#For Vehicle Recogniton for tiny-cnn net

1.What is Vehicle Recognition?

this project classify car according to car logo.Now the lasting Version just distinguish VM car from other.

2.Where from Training Data ?

The major sources of our dataset include images captured by ourselves,Medialab LPR Dataset [1].

3.this models?

You can get it in JNi/test.cpp file.

Code:

  static const bool tbl[] = {
			O, X, O, O, O, O, O, O, O, O, O, X, O, O, O, O,
			O, X, X, X, O, O, O, X, X, O, O, O, X, X, O, O,
			O, O, X, X, X, O, O, O, X, X, O, O, X, X, X, O,
			O, O, O, X, X, X, O, O, O, X, X, O, X, X, O, O,
			O, O, O, O, X, X, O, O, O, O, X, X, O, X, X, O,
			O, X, O, O, O, X, X, O, O, O, O, X, O, O, X, O,
	};

	 nn << convolutional_layer<tan_h>(40 , 40 , 3 , 1 , 6)  
		<< average_pooling_layer<tan_h>(38 , 38 , 6 , 2)   
		<< convolutional_layer<tan_h>(19 , 19 , 4 , 6 , 16 ,
		connection_table(tbl, 6, 16))              
		<< average_pooling_layer<tan_h>(16 , 16 , 16 , 2)  
		<< convolutional_layer<tan_h>(8 , 8 , 3 , 16, 16) 
		<< fully_connected_layer<tan_h>(16 * 6 * 6 , 64)
		<< fully_connected_layer<relu>(64 , 2);
		

My models have three conv-layers and two pooling layer , fully-connect layer , in the end layer , the activation functions is relu and the size of all conv-kernel is 3x3.The optimization algorithm of cnn is stochastic gradient levenberg marquardt.

Other arguments can't be public.

4.Experiment

data：

	 500 train image , 298 test image.

platform:

	Windows+VS2013.
Result:
	 the recognition rate is above 94.29% 

#How to use it to recognize face or other object?

If want to recognition other object , you must learning Cnn and tiny-cnn , contructing optimal model and training it using enough object images to get the wb-file(weights and bias values).Finish , replace /assets/tinyfile/carlogo file with wb-file.

#references

[1]Medialab LPR dataset, March. 2013[online]. Available: http://www.medialab.ntua.gr/research/LPRdataset.html

[2]Humayun Karim Sulehria, Ye Zhang.Vehicle Logo Recognition Using Mathematical Morphology

[3]Humayun Karim Sulehria, Ye Zhang.Vehicle Logo Recognition Based on Bag-of-Words.IEEE International Conference on Advanced Video and Signal Based Surveillance,2013.

#Running Screenshot

(1)For Vehicle Recogniton

(3)For gender Recognition

#Discussing

• issue
• email:[e-mail]