Multitargets Joint Training Lightweight Model for Object Detection of Substation.

Area of Science:

• Computer Vision
• Electrical Engineering
• Artificial Intelligence

Background:

• Traditional substation object detection struggles with diverse object classes and lacks generalization.
• Deep learning models offer better generalization but are computationally intensive, limiting their use in resource-constrained substation monitoring.

Purpose of the Study:

• To develop a lightweight object detection model for substations that overcomes the limitations of traditional and complex deep learning methods.
• To improve the accuracy and efficiency of object detection in substation environments.

Main Methods:

• Proposed a multi-targets joint training lightweight model using feature maps from complex models and object labels.
• Introduced the 'heat pixels' method to enhance object information, addressing foreground-background imbalance.
• Implemented and evaluated three lightweight networks trained with the proposed method on public (VOC) and custom substation datasets.

Main Results:

• The proposed model significantly improved object detection accuracy in substations.
• The method effectively reduced the time and computational resources required for object detection.
• Feature maps from complex networks provided deeper insights and higher information entropy for training.

Conclusions:

• The developed lightweight model offers a practical solution for substation object detection, balancing accuracy and computational efficiency.
• The heat pixels method is effective in improving object detection performance, especially in scenarios with imbalanced data.
• This approach enables the deployment of advanced object detection capabilities on substation monitoring terminals with limited computing power.