Improving YOLOv7's Remote Sensing Image Object Detection

In order to address the challenge of poor detection accuracy concerning both dense and smaller targets within remote sensing images set against complex backgrounds, we propose an algorithm for target detection within remote sensing images utilizing an enhanced YOLOv7 framework. Initially, the ELAN module within the backbone segment is replaced with the cascade fusion Network, aiming to more effectively leverage the parameters within the backbone network. This augmentation enhances the model's aptitude for multi-scale feature representation. Subsequently, the EMA attention mechanism is introduced and integrated into the prediction location, serving to encode global contextual information and recalibrate channel weights within the features. This adjustment notably amplifies the feature representation capability. Experiments were conducted on the DIOR and NWPU VHR-10 datasets, in comparison with the YOLOv7 algorithm, the mAP exhibited improvements of 2.6% and 2.2% on these respective datasets. Moreover, detection accuracies were enhanced compared to several existing mainstream algorithms. The outcomes underscore the improved algorithm's effectiveness in substantially augmenting the detection accuracy of remote sensing image targets within complex backgrounds and diminishing instances of leakage and false detections in remote sensing targets.