Abstract: Aiming at the problem that deep learning-based transmission line hardware detection methods can not ensure the lightweight model while improving the detection accuracy of small targets, a lightweight detection method for small target hardware on transmission lines is proposed. Based on YOLOv4, the method firstly combines density-based spatial clustering of applications with noise and the anchor frame optimization strategy with multiple K-means to optimize the preset frame selection according to the characteristics of the small target data, speed up the network convergence and improve the detection accuracy. Secondly, GhostnetV2 is used to achieve model light weighting. Then the network of small object enhanced multi-scale detection network, is designed to enhance the feature extraction ability, strengthen the fusion of shallow and deep features, and solve the problem of feature loss caused by down sampling cross-step convolution. Finally, focal loss function is used to optimize sample allocation and depth-wise separable convolution is used to reduce model complexity. The mean average precision(mAP) of this algorithm on the self-constructed small hardwaredataset reaches 64.73%, and the model computation and parameter counts are reduced by 78% and 71% compared to the pre-improvement period, while there is almost no loss of detection accuracy. The mAP reaches 38.43% on the public dataset VisDrone2019, which gives the algorithm a superiorsmall target detection performance compared to other algorithms.

Key words: transmission line hardware, density clustering, small target detection, light-weighting, feature fusion, focal loss function