Clothing image classification model using regularized multiple convolutional neural networks (RMCNN)

Abstract

In the past few decades, machines have gradually taken over the daily activities of human beings such as online shopping and clothes manipulation. It is essential to develop artificial intelligence techniques that can help people detect and classify clothing designs accordingly. Early efforts to solve the clothing image classification problem require carefully selecting and extracting certain features from clothing image datasets in such a way that the features of the datasets are highly represented. However, these methods are difficult in defining and capturing a wide range of image features. Research shows that Convolutional Neural Networks (CNN) models can solve image classification problems better than traditional machine learning (ML) methods. However, they are faced with problems such as over-fitting, Hyper-parameter tuning, Noisy data, and insufficient training data. This work addresses the problem of overfitting which reduces the classification/generalization performance of clothing image classification models. We proposed four (4) CNN models in which a Regularization method called Dropout is added to each layer to handle the over-fitting problem. The model with the best result out of the four is adopted as the proposed model. The results show about 1.77% improvement in accuracy as compared to the results recorded by other models that were trained using the same dataset and the state-of-the-art architectural designs.