Deep learning-based infrared target detection

Computer technology is developing gradually and is used in many fields today. One of the most important of these areas is computer vision. Thanks to advanced computer vision applications, object detection studies are carried out from images. Object detection studies, which are generally done on colour images, have started to work on infrared images with the development of thermal sensor technology. Infrared images are widely used in many areas, especially in defence technologies, medicine, and daily life. The use of infrared images provides a great advantage in object detection studies carried out in bad lighting and weather conditions. With the gradual development of deep learning architectures, Convolutional Neural Networks have been widely used in object detection. The YOLO models are the most widely used Convolutional Neural Network models in the literature, which stand out with their object detection speed and accuracy. In this study, people and vehicles were detected from infrared images to be used in defence technologies. In the scope of the study, infrared object detection performances and thermal domain adaptations of YOLOv3, YOLOv4 and YOLOv7 models were compared. In this context, a dataset consisting of 48000 images obtained by data augmentation methods was used. The dataset was given to the models with the 4-fold cross-validation method. Each model has trained 100 epochs and the results were evaluated. In the experimental results, it was observed that the YOLOv3 and YOLOv4 models had the best accuracy. The mAP of the models were obtained as 84.93%, 95.97% and 57.29% for YOLOv3, YOLOv4 and YOLOv7, respectively.