A Lightweight Multi-label Image Classification Model Based on Inception Module

Abstract

Convolutional Neural Networks (CNNs) have shown enormous potential for solving multi-label image classification problems. In recent years, a lot of experimentation is done with various state-of-the-art CNN architectures. The CNN architectures have evolved to become deeper and more complex in these years. These architectures are big due to a greater number of layers and trainable parameters. However, there are many real-time applications which demand fast and accurate classification. Keeping this in consideration, a simple model inspired by Inception V7 is proposed for multi-label image classification in this work. The proposed model consists of six convolution layers including three inception blocks with one million parameters approximately, which are very few as compared to many state-of-the-art CNN models. This makes the model deployable in lightweight devices for some real-time applications. The comparison experiments with other deep state-of-the-art CNNs were carried out on image datasets from multiple domains including general benchmark datasets, medical datasets, and agricultural datasets. The model exhibits better performance on many datasets making it feasible to use in various domains for multi-label image classification.