Image Database Generation and Management System for the Small-pixelized Airborne Target Recognition

Abstract

Recently, the number of cases of unauthorized intrusion into flight prohibited areas within airport control is increasing rapidly. Most hobby drones are small objects of less than 2kg, and image-based object tracking systems are greatly affected by detection accuracy depending on the small size and image resolution of the drone. if an object is more than 200m away, it is difficult to identify existing object characteristics with small objects of 32x32 pixels or less. In this study, it was confirmed that micro-objects of 32x32 pixels or less can be effectively extracted by histogram binaries based on objects in images using the classical image processing method of HOG (Histogram of oriented gradients) and LBP (Local Binary Pattern) and to identify micro-objects of airborne target datasets collected using handmade tools, pre-processed image dataset is constructed by applying classical image processing techniques, and YOLOv5-based learning is conducted to improve object detection accuracy of 32x32 pixels or less. In addition, there are not many public image datasets (Datasets) for drones and airborne target in flight, we collected images including airplanes, drones, and birds. Qualitative data augmentation was performed using classical image processing methods using the collected datasets, classified into four datasets, and then the YOLOv5s model-based inference performance was analyzed. As a result of the analysis, it was confirmed that the inference performance of the dataset that performed qualitative augmentation using the classical image processing method was improved.