A Study on Performance Evaluation of MWIR Image Detection Based on YOLO Model Using RGB Channel Image

Abstract

Recently, artificial intelligence is being used in many business fields. In the field of image, it is used in many different forms, starting with simple object detection, tracking, synthetic image generation, and style conversion. In particular, the object detection field has already been applied and used in many fields such as national defense, product defect detection, and security thanks to tremendous development. However, current object detection models are mainly performed with RGB images. Due to this direction of research, a separate study is underway for a model for IR image. Because of this, the development of deep learning models for IR images is much slower than RGB images. In addition, due to the lack of IR image data, research on IR image deep learning models is becoming more and more laggy compared to other deep learning studies. This paper proposes that the model trained on RGB images shows excellent performance in IR images. The object detection deep learning model learns shape information by using feature extraction. Our results show that IR images showing the shape of an object and images learned as RGB images can be sufficiently inferred. As a result, the model trained with RGB images shows robustness even in IR images.