Aiming at the problems of low detection accuracy of power distribution equipment components in complex inspection environment and the difficulty of deploying algorithms to mobile terminals such as UAVs

we put forward a combined pruning and knowledge distillation method for the detection of distribution line equipment components. Firstly

the GSConv Convolution is introduced into the feature fusion network instead of the Standard Convolution to fully utilize the feature information extracted from the backbone network; then

the Context Guided Block is fused into the feature extraction network to improve the accuracy of equipment components detection. Secondly

Inner-CIoU loss function is introduced into the detection head to enhance the regression effect of low IoU samples. Finally

the model is compressed using the joint pruning and knowledge distillation methods to reduce the number of parameters

floating-point operations

model size and increase the speed of detection. In order to validate the effectiveness of the proposed method

a distribution line components equipment dataset is constructed by using UAV inspection images for testing. The test results show that

compared with the benchmark network

the designed detection method can be adopted to improve the detection accuracy mAP0.5 and mAP0.5:0.95 by 4.7% and 2.3%

respectively; meanwhile

the model size is reduced by 70.6%

the number of parameters and floating-point operations is reduced by 71.3% and 46.8%

respectively

and the detection speed can be up to 54.4 frames per second

which can satisfy the real-time requirements of the UAV inspections on distribution lines.