Research Paper Publication:

SangCTC- Improving Traditional CTC Loss Function for Optical Music Recognition (https://doi.org/10.1109/PuneCon46936.2019.9105720)

Citation:

J. K. Shah, A. S. Padte and P. N. Ahirao,
"SangCTC-Improving Traditional CTC Loss Function for Optical Music Recognition (August 2019),"
2019 IEEE Pune Section International Conference (PuneCon), Pune, India, 2019, pp. 1-5,
doi: 10.1109/PuneCon46936.2019.9105720.

Focal-CTC-OMR

(Implementation of Part I of SangCTC)

Problem Statement

Given a music sheet, usually in the form of an image, the goal of an OMR system is to use various vision algorithms to interpret the corresponding music symbols and later convert it into digital playable format of music. To solve this problem in an End-to-End manner, Convolutional Recurrent Neural Network (CRNN) architecture is used. It considers both spatial and sequential nature of this problem. CTC loss function is proved to be a favorable choice in these types of sequence problems as it trains the models directly from input images to their corresponding musical transcripts without the need for a frame-by-frame alignment between the image and the ground-truth thereby solving the purpose of End to-End training. Though traditional CTC seems to solve a major chunk of the problem, it suffers from some limitations due to Unbalanced Dataset.

python

Unbalanced Dataset Problem

Data Visualization

python
python
python

Solution: Focal CTC

Focal CTC Loss Function

def focal_ctc(alpha=0.5,gamma=2.0,targets,logits,seq_len):

    #FOCAL LOSS
    #This function computes Focal Loss
    #Inputs: alpha, gamma, targets, logits, seq_len
    #Default Values: alpha=0.5 and gamma=2.0
    #Output: loss

    ctc_loss = tf.nn.ctc_loss(labels=targets, inputs=logits, sequence_length=seq_len, time_major=True)
    p= tf.exp(-ctc_loss)
    focal_ctc_loss= tf.multiply(tf.multiply(alpha,tf.pow((1-p),gamma)),ctc_loss) #((alpha)*((1-p)**gamma)*(ctc_loss))
    loss = tf.reduce_mean(focal_ctc_loss)

return loss

Edit: If you are facing Exploding Gradients Problem ('nan' in loss) try clipping on ctc_loss

Experiments and Results

python
python

Sampling the Dataset

(Scripts contain sampling and visualization codes)
python

Word Cloud