RSODNet: Lightweight Remote Sensing Image Object Detection Combined with BCDNS Compression Algorithm

In recent years, with the gradual increase of neural network Params (the aggregate of trainable elements in a model, including weights, biases, and other adjustable elements) and calculation volume, model compression within an acceptable range of network accuracy variations has emerged as a prominent research focus in the field of deep learning. Model pruning and knowledge distillation have been widely used for reducing the complexity and storage cost of neural networks. This study designs the Remote Sensing Object Detection Network (RSODNet), a lightweight model for remote sensing image object detection, and proposes bridging cross-task distillation network slimming (BCDNS) as a method that integrates model pruning and knowledge distillation. The experiment results indicate that RSODNet outperforms the YOLOv8 model in various metrics while maintaining almost unchanged Params and calculation volume. The BCDNS method eliminates redundant channels while preserving a priori knowledge of the initial model intact. This study offers technical support for compressed models used in object detection from remote sensing images