synthetic image feature detection using deep learning

Aim:

to implement CNN based feature detector to detect corners in synthetic images genrated using opencv or some other method. the some sample images used are shown below.

sample input images used for training and testing

Generating label using opencv:

1. Import datset to train
2. Generating label Images:
   • label image were generated using opencv "Harris corner detector".
   • the corner point detected by harris corner detector were replaced by gaussian kernel.
   • 
   • thus the generated image is been used for training the network.
   • image and label pair generated using above method:
   

network architecture:

Network Consist of encoder and decoder both constructed using structure similar to resnet's residual block which takes images as input and generate corner feature map as output shown below.

results:

hyperparameters:

Parameters	values
learning_rate	0.001
momentum	0.9
epoch	1
batchsize	128

Conclusion:

the network was successful in detecting feature-points given a raw image.