DEEP LEARNING IMAGE RECOGNITION WITH INTELLIGENT DIMENSION REDUCTION

Say you want to write a program which is supposed to recognize a certain type of truck. But as a resource, you only have about 10 thousand annotated images avaible. So, this would not be enough data for fully training a resnet 54 for example. Hence, you use pretrained parameters provided by google, which has trained the model on billions of images. Then, you just train the last layer using your 10 thousand images. This approach can have many practical draw-backs, for example that it learns all kind of noise associated with the specific collection of your small data-set. We developped an algorithm, which can avoid many such problems, by first understanding in an intelligent way the rough features of the object to be recognized. Here is

part of the script of intelligent preprocessing

If you would be interested in getting the complete algorithm, you can E-mail us.