Multi-scale object detection network based on selective large convolution kernel

Deep learning has developed various object detection networks for different scenarios, but remote sensing imagery presents challenges due to its top-down perspective, causing issues with target orientation, scale variation, and obstructions. To tackle these, we introduce LSMA-YOLOv8, which enhances YOLOv8's speed and accuracy. It uses a backbone with large convolution kernels for better feature extraction and a multi-scale module for detecting targets at various scales, which is disabled during inference to reduce computation. LSMA-YOLOv8 also includes YOLOv7's dual detection heads for optimized performance, offering an efficient solution for object detection in remote sensing images.