A Multiple-Loss Dual-Output Convolutional Neural Network for Fashion Class Classification
Okeke Stephen, Uchenna Joseph Maduh, Sanjar Ibrokhimov, Kueh Lee Hui, Ahmed Abdulhakim Al-Absi, Mangal Sain · 2019
An improved multi-loss multi-output convolutional neural network method was deployed to extract features from a given set of disjointed data (Fashion and Color) with diverse convolutional chunks in a single network. The first convolution block extracts features from the first image dataset (Fashion) and determines the classes to which they belong. The second block is responsible for learning the information encoded in the second set of data (color), classify and append such to the features extracted from the first convolutional block. Each block possesses its loss function which makes the network a multi-loss convolutional neural network. A set of double fully connected output heads are generated at the network terminal; enabling the network to perform predictions on a combination of disjointed labels. To validate the classification ability of our network model, we conducted several experiments with different network parameters and variations of data sizes and obtained remarkable classification results of 98 and 95 on the fashion and color sets respectively.