Multi neural network model for product recognition and labelling

Deep Convolutional Networks have demonstrated very good performance on classifying the ImageNet dataset. This work introduces a deep learning model with a unique architecture that performs classification on a custom dataset built from scratch. We report an architecture that consists of a network of classifiers, which incorporates the state of the art feedforward network architecture. Our system is built by placing a master classifier that forwards the image to 3 of the 10 classifiers that in turn perform the classification. We have trained the system for 131 different classes comprising of 1920 images of commonly found grocery products and commodities. On testing we received a top-1 error rate of 38.62 percentage, top-3 error rate of 25.36 percentage and a top-5 error rate of 20.28 percentage. Each of the 10 classifiers have been trained independently on a subset of the 131 classes. The master classifier is trained on the combined dataset. We use Transfer Learning and retrain the Inception v3 model for each classifier.