LW-YOLOv7-Tiny for Remote Sensing Image Object Detection

Real-time detection of targets in remote sensing images is a key technology for efficient applications in the field of remote sensing. Object detection models based on deep neural networks have high detection rates, but such algorithms often have many parameters and large computational requirements, making them difficult to deploy on edge devices. A lightweight remote sensing image object detection method, LW- YOLOv7-tiny, is proposed for the aforementioned problem. The MobileNetV3 network is introduced into the YOLOv7-tiny model, the feature extraction network is optimized, and the ability to extract features from images is enhanced, resulting in richer semantic information. The method reduces computational complexity while maintaining high detection accuracy, achieving lightweight optimization of performance. ECA-Net is introduced at the output of the backbone network to weaken the expression of non-attentive points while enhancing the information channel of key points, further improving detection accuracy. The SiL U activation function is used to improve the performance and generalization ability of the model and prevent overfitting. To verify the effectiveness of the model, testing was conducted on the RSOD dataset and compared with existing mainstream algorithms. Experimental results show that the improved model has an mAP improvement of 1.68, with a reduction in parameter volume and computational complexity of 0.15M and 0.62G respectively compared to the original model. The improved LW-YOLOv7-tiny achieves high accuracy while lowering the model's parameter volume and computational complexity, further improving detection results, providing feasibility for real-time detection of remote sensing images and deployment on edge devices.