YOLO OPTIMIZATION FOR EDGE AI: A LIGHTWEIGHT APPROACH FOR DEEP SKY OBJECT DETECTION

Accurate and efficient object detectors on resource-constrained edge devices are essential for real- time space applications such as satellite monitoring and deep sky exploration. However, Deep Learning (DL) models face difficulties when implemented in environments with limited memory and computational capabilities. To address this problem, this study proposes two lightweight architectures based on YOLOv10s and YOLOv10n that integrate ShuffleNetv2 and MobileNetv2. The Shuffle architecture improves efficiency by taking the advantages of the depthwise separable convolution and channel shuffling. The Mobile-Shuffle architecture combines Shuffle blocks with MobileNetv2-inspired blocks to enhance feature extraction and representation while maintaining low computational cost. This work evaluates both architectures on the augmented DeepSapceYOLODataset dataset, which contains a collection of annotated deep sky objects with different sizes and complexities. The evaluations demonstrate that the Shuffle architecture achieves the highest mAP50 on both sets, the fastest inference speed, and the smallest model size, especially in YOLOv10s architecture. While the Mobile-Shuffle architecture shows its strength in the YOLOv10n variant by improving the detection accuracy, it comes at the cost of the increased inference latency. These findings demonstrate that the proposed models offer a practical trade-off between accuracy, speed, and size, making them well-suited for edge deployment in space-related scenarios.