Multi-Label Classification using Deep Convolutional Neural Network
A. Agnes Lydia, F. Sagayaraj Francis · 2020
Convolutional Neural Networks (CNN) have proven to perform well in single-label classification tasks. However, performing a multi-label classification using a Convolutional Neural Network is still under research. Several variants of Convolutional Neural Networks have come into existence due to extensive research work with numerous improvisations. From the several variants of CNN, VGG-Net has been proven to provide a stable performance and hence it is used for experimental purposes in this paper. Generally a network trained using a specific classifier will be tested using the same classifier, to test the learning capability of the model. Models trained in this manner has the capability to recognize only a single object at a time. To overcome this scenario, this work intends to train a VGG-Net to recognize more than one label in a single instance of image sample, without increasing the complexity of the network architecture. To implement this objective, the network is trained and tested using Sigmoid: a binary classifier, instead of using Softmax: a multi-class classifier. This proposed model is experimented on self-curated datasets scrapped from Google images and improvised using manual pruning for relevancy and balanced samples in each category. The first dataset created is Colour Clothes and the second dataset created is Fruits & Vegetables. The performance of the network is evaluated with standard metric, Binary Cross-Entropy.