From the perspective of aerial photography of UAVs(unmanned aerial vehicles) during power inspections, taking into account the particularity of the image characteristics of power equipment, a visual detection model EDR-YOLOv7 is proposed suitable for the edge of UAVs to address the common problems of the ubiquitous small target detection, target point occlusion, and the increase in model missed detection rate caused by the variable scale of aerial images, as well as increased calculation amount caused by dense detection. Firstly, a display visual center module is introduced into the neck network to capture the implicit relationship of pixels and solve the problem of missing small target features. Secondly, the dynamic sampling module is used to replace the transposed convolution to achieve flexible sampling of feature points and reduce the complexity of model calculation. Finally, in order to solve the problem of variable viewing angle scale of UAVs and the problem of accidental deletion and deviation of prediction frames caused by partial occlusion, the Inner-SIoU (inner-scylla intersection over union) loss term and the repulsion factor are added to the loss function, continuously reducing the prediction error during training iterations. After experimental verification, EDR-YOLOv7 compared to the original model increases mAP@0.5 and the detection frame rate by 3.89 % and 5.2 frames/s respectively. The model is finally deployed on the Jetson XAVIER NX edge computer and accelerated by TensorRT reasoning, which performs well in video stream detection tasks.