Implementing GoogLenet for Devanagri (Hindi) character classification.

[Harshit Laddha]

I want to implement Googlenet for hindi character classification. This net has been used to classify chinese characters in this research paper--  http://arxiv.org/pdf/1505.04925.pdf.

Till now I am only using Googlenet without embedding any directional feature maps.

I have used Googlenet with one and two inception modules. After running for 35 thousand iterations both networks are giving accuracy around 88.5% which almost remains constant till 60 thousand and starts declining  afterwards.

No. of classes = 60

Dataset size = 25000 (both test and train)

Image size = 64x64

base_lr: 0.001
momentum: 0.9
weight_decay: 0.0005
lr_policy: "inv"
gamma: 0.0001
power: 0.75

I just wanted to know is this optimum accuracy for these nets on this data or nets can be optimized further. This research paper has mentioned accuracy of 96.26% on Chinese Character which are far more in no. of classes and complex as compared to Hindi characters.
But on hindi characters accuracy is stuck at 88%. What can be done to further increase accuracy.

I am attaching dataset, solver and network architecture files for both networks.


Thanks

attachments-

Network architecture file for Googlenet with 1 inception module : google.protoxt

Solver for Googlenet with 1 inception module : google_solver.prototxt

Network architecture file for Googlenet with 2 inception module : google2.protoxt
Solver for Googlenet with 2 inception module : google2_solver.prototxt

 image data : hindi_64.zip

lmdb data : lmdb_data.zip

[Yang Feng]

Currently I am ddoing the same thing to recognize the handwritten chinese character,for my dataset,i have 4957 classes, and the training set is 237000, and the test set is 54000,and the input image is set by myself because my data is a series of location data. You share your data and lmdb, but i didn't find it. I will try your data,mybe you can contact me through this email  hit.ya...@gmail.com. We can discuss this together! I am looking forward for your reply!

[Miguel Garcia]

Try centering your training images