rhemon/food_rec

Neural Network for recognising food items from its picture.

food_rec

Training Set: (4800, 2000)

Items:

* Green Apple
* Mango
* Pear
* Orange

Accuracy Results:

2 Layer Neural Network
12 - 4
0.02 learing rate
3000 iterations

Accuracy with training set, 85.109375%
Accuracy with dev set, 84.75%
Accuracy in recognising mango: 78.5555555556
Accuracy in recognising apple: 93.6666666667
Accuracy in recognising orange: 94.3333333333
Accuracy in recognising pear: 64.4444444444

2 Layer Neural Network
4800 - 4
0.008 learning rate
2000 iterations

Accuracy with training set, 82.890625%
Accuracy with dev set, 82.0%
Accuracy in recognising mango: 74.5555555556
Accuracy in recognising apple: 88.8888888889
Accuracy in recognising orange: 91.4444444444
Accuracy in recognising pear: 67.0

3 Layer Neural Network
4800 - 2400 - 4
0.008 learning rate
2000 iterations

Accuracy with training set, 75.046875%
Accuracy with dev set, 78.5%
Accuracy in recognising mango: 71.8888888889
Accuracy in recognising apple: 86.0
Accuracy in recognising orange: 84.5555555556
Accuracy in recognising pear: 50.0

food_rec's Issues

Reducing bias

• longer networks

• larger input images

• better optimisation algorithm

Softmax for output layer

Use softmax in output layer to predict the item.