TL-DREN: Transfer Learning Based Detection and Recognition of Electricity Nameplates

This paper presents a novel approach that integrates traditional image processing techniques with deep learning to enhance the accuracy and robustness of fuzzy text detection and recognition in nameplate images of power equipment. Initially, traditional methods are employed to locate the nameplate and extract information, followed by perspective transformation to correct image distortions. Subsequently, deep learning models are utilized for text detection and recognition. Two specialized datasets were constructed for this study: one for text detection, comprising 841 high-resolution images captured under various environmental conditions, and another for text recognition, consisting of 42,556 cropped text region images derived from the corrected nameplate images. The PP-OCRv4 model was fine-tuned on our curated dataset to improve its performance on text detection and recognition tasks specific to power equipment nameplates. Experimental results demonstrate that our proposed method achieved an F1-score of 0.524 for text detection and an accuracy of 0.82 for text recognition, showcasing the potential of this approach to enhance the efficiency of automated management of power equipment.