Remote Sensing Small Target Detection Research Based on Improved YOLOv7

This study introduces a streamlined improvement algorithm based on YOLOv7 to address the challenges of multiple small targets, complex backgrounds, limited feature points, and environmental influences in remote sensing images. The algorithm incorporates the GSConv module and upsampling to achieve more effective feature fusion. The GSConv module combined with the VoV-GSCSP module achieves Slim-Neck for reducing algorithm complexity and improving accuracy. The Receptive Field Enhancement (RFE) module is introduced after the spatial pyramid pooling layer (SPPCSPC) to enhance the receptive field of small targets. By changing the settings of convolution operations and increasing the receptive field size of each convolutional layer, local and global features in the image are better captured. Information flow is maintained through skip connections. Experimental findings show that the enhanced YOLOv7 algorithm reaches an mAP of 95.5% on the NWPU VHR-10 dataset, marking a 6.4% improvement over the original algorithm’s mAP. That improves the accuracy of object detection in remote sensing images.