Modified YoloV10x for Object Detection on Drone-based Image Sensing

Images captured by an AI-powered drone are often used for various purposes, including object detection. However, detecting small objects remains challenging in drone-captured images, requiring a reliable deep-learning architecture to handle feature extraction and image downsampling during the machine-learning process. This research aims to propose a deep-learning architecture to tackle such challenges. Based on the YoloV10x architecture, the conditional identity blocks in the backbone part are replaced by convolution and bottleneck blocks for faster and more efficient downsampling. Using the VISDRONE DET dataset, the learning process was performed from scratch. The model's performance evaluation uses mean Average Precision (mAP) and loss functions. Compared to the baseline architecture called Realtime Detection Transformer (RT-DETR), the proposed architecture can increase by approx-imately an average of 27% mAP scores for each class in the dataset.