A Novel Method for Improving Baggage Classification Using a Hyper Model of Fusion of DenseNet-161 and EfficientNet-B5

Abstract

In response to rising concerns over crime rates, there has been an increasing demand for automated video surveillance systems that are capable of detecting human activities involving carried objects. This paper proposes a hyper-model ensemble to classify humans carrying baggage based on the type of bags they are carrying. The Fastai framework is leveraged for its computational prowess, user-friendly workflow, and effective data-cleansing capabilities. The PETA dataset is utilized and automatically re-annotated into five classes based on the baggage type, including Carrying Backpack, Carrying Luggage Case, Carrying Messenger Bag, Carrying Nothing, and Carrying Other. The classification task employs two pretrained models, DenseNet-161 and EfficientNet-B5, with a hyper-model ensemble that combines them to enhance accuracy. A “fit-one-cycle” strategy was implemented to reduce the training time and improve accuracy. The proposed hyper-model ensemble has been experimentally validated and compared to existing methods, demonstrating an accuracy of 98.6% that exceeds current approaches in terms of accuracy, macro-F1, and micro-F1. DenseNet-161 and EfficientNet-B5 have achieved accuracy rates of 95.5% and 97.3%, respectively. These findings contribute to expanding research on automated video surveillance systems, and the proposed model holds promise for further development and use in diverse applications.

Keywords

• deep learning
• hyper model
• pre-processing pipeline
• baggage type
• Fastai framework