Comparative Study of CNNs for Camouflaged Object Detection

Convolutional neural networks (CNNs) have revolutionized the computer vision field, achieving remarkable performance in various visual recognition tasks, including object detection. One challenging problem in object detection is the detection of camouflaged objects in natural scenes. Camouflaged object detection (COD) is an essential task in various real-world applications, such as military surveillance, autonomous driving, and robotics. In recent years, numerous methods have been proposed for this task. The existing research on this issue has relied either on biological traits of the camouflaged objects or computer vision methods. In recent years, significant progress has been made in this field, with various techniques proposed for the detection of camouflaged objects. The detection and tracking of camouflaged objects using computer vision algorithms is discussed in this article from a theoretical standpoint. The proposed work presents a comprehensive review of the different methodologies used for camouflaged object detection, which includes an overview of the most recent cutting-edge techniques for COD and evaluate their performance on three popular Benchmark Datasets: COD10K, CAMO, and Chameleon. We analyse five different models, including SINet, SINet-V2, DGNet, ERRNet, and HITNet, based on two evaluation metrics: S-measure and E-measure. After evaluating the approaches in a common setting of dataset and computing system, the result is provided. Our results demonstrate that HITNet outperforms the other models across all datasets with an accuracy of 97%. Therefore, HITNet model underwent post-dynamic range quantization to reduce its size from 101 MB to 25.5 MB to enable deployment on edge device.