Application of YOLOv7 in Remote Sensing Image Target Detection

Remote sensing images are characterized by small targets, complex types, and large variations between different and similar targets, resulting in poor target detection and uneven detection performance of multiple types of targets. To reduce the error detection and omission detection of remote sensing image targets and improve the detection accuracy of multiple targets, an improved feature extraction network YOLOv7 detection model is proposed. Based on the YOLOv7 algorithm, the model was experimentally studied using the DOTA remote sensing image dataset. A new ELAN module is constructed using full-dimensional dynamic convolution ODConv, which improves the utilization of multi-dimensional information in nuclear space and reduces computation volume. Integrates the multi-headed self-attentive MHSA into the backbone network to better distinguish background information. Our proposed network model has 4.6%, 15.9%, 4.1% and 3.7% mAP improvement compared to YOLOV3-CSP, YOLOV4-CSP, Resnet50 and X50-CSP, respectively. Most importantly, Compared to the original YOLOv7 algorithm, the parameter settings are reduced by 6.6%, and the average detection accuracy(mAP) improved by 1.44%, better balance between accuracy and the number of parameters while meeting real-time detection.