Improved Gaussian Distance Loss for Photovoltaic String Rotation Object Detection Algorithm

In response to the problems of low detection accuracy and inability to accurately locate photovoltaic strings caused by complex backgrounds, dense target distribution, and diverse target directions during the current unmanned aerial vehicle inspection of photovoltaic strings, an improved Gaussian distance loss photovoltaic string rotation target detection model is proposed based on the single stage target detection network RetinaNet. In the feature extraction stage, a backbone network ResNest-G integrating channel attention and random regularization nonlinear activation function GELU was designed to enhance the extraction of photovoltaic string edge features. The GELU activation function was introduced to improve the convergence speed of the training process; And introduce Gaussian Wasserstein Distance (GWD) as the regression branch loss function, optimize the model to learn the rotation angle information of the photovoltaic string detection frame to achieve precise positioning. The experimental results show that the improved model can achieve mAP of 84.1% on the self-made photovoltaic string dataset, which is 7.8% higher than the original RetinaNet network model. It can accurately identify and locate photovoltaic panels, proving the effectiveness of the improved model in photovoltaic string target detection.