A Novel Obejct Detection Framework for Power Equipment Recognition in Pumped Storage Power Station

Power equipment recognition in pumped storage power station is still a challenge due to complex environment and limited training samples. To solve this problem, we introduce several techniques including spatial sampling, local descriptor extraction and balanced learning. Spatial sampling considers the relationships between spatial features in the map and enhances the feature maps with the second-order information. To focus on certain type of power equipment, we extract local descriptors which are invariant to transformations. Finally, to solve the imbalance of classes of power equipment, we adopt balanced learning which includes balanced Intersection over Union (IoU), balanced feature pyramid and logarithmic gradient re-weighting. Balanced IoU tackles imbalances in the dataset by prioritizing hard samples. Balanced feature pyramid ensures that features from different network levels are equally represented. Logarithmic gradient re-weighting balances the gradients during training. Experiments on our inspection image dataset show that the proposed method is much better than baselines while improving accuracy of small equipment and reducing the risk of overfitting.