Machine Learning Projects for Mobile Applications

Distortions

We can improve the results by giving tough input images during training. Training images can be generated by cropping, brightening, and deforming the input images randomly. This will help in generating an effective training dataset. However, there is a disadvantage of enabling distortion here, since bottleneck caching is not useful. Consequently, the input images are not reused, increasing the training time period. There are multiple ways to enable distortion as shown here:

--random_crop--random_scale--random_brightness

This won't be useful in all cases. For example, it won't be helpful in a digit classifier system, since flipping and distorting the image won't make sense when it comes to producing a possible output.