Abstract: Aerial images of transmission line fittings have the characteristics of small target proportion and excessively complex background. In order to transfer the performance of a teacher network which has excellent performance in detecting fitting targets to a student network, a decoupled knowledge distillation method based on shared feature scoring for transmission line fitting detection is proposed. Firstly, the image features are decoupled into foreground and background regions according to the Ground-truth box. In the foreground region, the feature information of each layer in the teacher model's feature pyramid is classified and scored, and the classification scores of all categories are aggregated as the common feature distillation mask. Secondly, in order not to maintain the integrity of the whole image, the GcBlock module is used to capture the relationship among the object of the fittings, the background and other fittings, so as to transfer the image feature knowledge generated by the teacher model to the student network completely. Finally, the validity of the proposed method is verified by a self-built fitting image detection data set. The experimental results show that, by applying this method to both two-stage and single-stage models, the detection performance of the network with a small number of participants can be greatly improved. The mean average precision of student network at 0.5 intersection over union of Faster R-CNN and RetinaNet can be improved by 25.9% and 31.4%, respectively, surpassing even the detection accuracy of the teacher network. The cascade R-CNN's student network has significantly improved targets for the shock hammer, adjustment board, and weight. The method in this paper can be adopted to realize the high efficiency detection of the fitting target, and to achieve the balance between detection performance and resource consumption.