Exposure image correction of electrical equipment nameplate based on the LMPEC algorithm

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Electrical equipment nameplate recognition is crucial for maintenance and inventory.
• Overexposure and underexposure in nameplate images hinder accurate data extraction.
• Existing algorithms struggle with poor image quality, necessitating improved methods.

Purpose of the Study:

• To develop an optimized algorithm for rectifying nameplate image exposure.
• To enhance the accuracy and robustness of electrical equipment nameplate recognition.
• To improve upon the LMPEC algorithm for image quality enhancement.

Main Methods:

• An optimization algorithm based on the LMPEC algorithm was developed.
• A PS-UNet++ network utilizing sub-pixel convolution was employed for feature extraction.
• Smooth L1 loss was used in the loss function to prevent model oscillation.
• Multi-scale training was incorporated to increase model robustness.

Main Results:

• The optimized algorithm demonstrated superior performance on a generated dataset of electrical equipment nameplate images.
• Significant improvements were observed in image quality metrics: SSIM (+5.6%), PSNR (+5.1%), and PI (+7.96%) compared to the original LMPEC algorithm.
• The enhanced algorithm effectively addressed overexposure and underexposure issues.

Conclusions:

• The proposed optimized algorithm effectively rectifies nameplate image quality, enabling more reliable recognition.
• The integration of PS-UNet++ and modified loss functions contributes to improved feature extraction and model stability.
• This approach offers a robust solution for enhancing the discoverability and usability of data from electrical equipment nameplates.