Deep Convolution Neural Network sharing for the multi-label images classification

Abstract

Addressing issues related to multi-label classification is relevant in many fields of applications. In this work. We present a multi-label classification architecture based on Multi-Branch Neural Network Model (MBNN) that permits the network to encode data from multiple semi-parallel subnetworks or layers outputs separately. Different types of neural networks can be used in the MBNN, but the proposal is made with Convolutional Neural Networks subnetworks, trained, and joined in classifying the outputs (i.e., labels). The proposed work makes it possible to perform incremental changes on existing Multitask Learning architectures for an adaptation to the multi-label classification. These transformations lead us to define two new architectures (neural network multi-outputs and neural network multi-features) using the feature extractors from the pre-trained neural networks. The empirical and statistical results verify that the proposed multibranch neural network architecture performs better than other simple multi-label classification architectures. Later, the “network with multi-features” obtained the highest classification score than other deep neural networks with 83.31% of the f1-score for the Amazon rainforest dataset. The f1-score values are 88.81% for Pascal VOC 2007 dataset, 87.71% for Nuswide, and 88.64% for Pascal VOC 2012.

Keywords

• Images multi-label classification
• Multi-branch neural network
• Deep convolutional neural network
• Multitask learning
• Transfer learning