Multi-layer capsule network with joint dynamic routing for fire recognition

Abstract

Fire recognition has emerged as a critical concern in the field of fire safety. Deep learning techniques, specifically convolutional neural networks (CNNs), have found widespread application in fire recognition tasks. The capsule network, a higher-level variant of CNN, offers distinct advantages in terms of enhanced recognition accuracy, making it suitable for fire recognition applications. However, the capsule network faces challenges in effectively determining the presence or absence of fire objects due to the idiosyncrasies of its dynamic routing algorithm. To address this issue and enable effective fire recognition, we propose a novel approach that involves a multi-layer capsule network. Within this framework, we introduce a joint dynamic routing algorithm that computes output values during forward propagation within the multi-layer capsule network. Additionally, we present a new loss function and a fully connected auxiliary training layer designed to train the multi-layer capsule networks effectively. Comparative evaluations against conventional CNN architectures and recent state-of-the-art fire recognition methods demonstrate that the enhanced multi-layer capsule network achieves higher test accuracy, even with limited training samples and fewer training iterations. Furthermore, owing to its reduced model parameter scale, the multi-layer capsule network exhibits faster recognition speeds compared to the aforementioned methods.