Object-Based Scene Classification Modeled by Hidden Markov Models Architecture

1. Introduction

Having the knowledge of surrounding environments is a major advantage for any existing agent (human or robot) in taking decisions or achieving tasks. Therefore, the ability to classify the current scene into a specific label guides the agent into accomplishing a better work and meeting expectations. However, the state of the art of scene classification reveals huge and persistent difficulties in implementing a reliable classifier. The accuracy in classifying a scene and the number of scene categories (number of classes) remain major aspects in determining the classification consistency. Therefore, scene classification problem became an open and a challenging area of research. This paper addresses the scene classification problem (Li, 2010) (Sikirić, 2014) with objects as attributes shortened as (SC:O) modeled by hidden Markov models (HMM) architectures and algorithms (Ghahramani, 2001). This approach was chosen since the HMMs are well known to be strong and reliable mathematical tools for classification and prediction. Moreover, HMMs treated efficiently and with great success similar problems such as speech recognition (Gales, 2008) (Gautam, 2017), speech synthesis (Reddy, 2017), machine translation (Wang, 2017) (Vogel, 1996), handwriting recognition (Sangeetha, 2017), activity recognition (Ozawa, 2017) (Alp, 2017), sign language recognition (Khandelwal, 2017) etc. Following the same perspectives, scene classification problem is a very active and attractive area of research. It is found in several research domains such as traffic road (Sikirić, 2014) (Lin, 2011) where it helps in taking decisions and organizes traffic, streets and airports scene surveillance systems (Lin, 2007) (Foresti, 1998) (Besada, 2001) where the classification process suggests and points out anomalies and suspicious behaviors. The scene classification can as well be found in area of research treating navigations (Liu, 2019) (Chen, 2019) where robots benefit of the semantic information about the surrounding environment provided by the category of scene. Several other types of scenes can be addressed such as areal scenes (Zheng, 2019) (Devi, 2019), indoor scenes (Li, 2019) (Hayat, 2016), outdoor scenes (Payne, 2005), or even war scenes (Raja, 2012). It can also be benefit to prediction systems where, in some circumstances, actions are predicted from a given scene categories (Vu, 2014).