Title:	2CNN - MODULAR IMAGE RECOGNIZER WITH CONVOLUTIONAL NEURAL NETWORKS
Author(s):	Kamil Szyc and Henryk Maciejewski
ISBN:	978-989-8533-80-7
Editors:	Ajith P. Abraham, Jörg Roth and Guo Chao Peng
Year:	2018
Edition:	Single
Keywords:	CNN, Modular, Re-Training Model
Type:	Full Paper
First Page:	11
Last Page:	18
Language:	English
Cover:
Full Contents:	click to dowload
Paper Abstract:	Deep convolution neural networks (CNN) have demonstrated remarkable performance on visual recognition tasks. This performance is feasible due to availability of massive datasets used for training of the networks, and requires long training time even on fast GPUs. One of the challenges related to application of CNN for image recognition is training a new class/category into the model. This usually requires that training data for the already learned classes is available to be used for re-training the model with the training examples for the new class. In this work we propose a modular architecture of image recognition based on n CNNs trained to recognize n individual image categories rather than train one CNN with n-way Soft Max as the last layer. We empirically show that this approach offers comparable performance with state-of-the-art CNNs, while allowing for much easier expansion in terms of classes learned by the model as compared with standard models based on a single CNN with an n-way Soft Max.
	Go Back