Transformer Image Quality Assessment Based on Multi-directional Feature Extraction

Wenqing Zhao, Mengwei Li, Yumei Ma

Abstract: Transformer image quality is influenced by various factors, including acquisition equipment and external environmental conditions. There is an urgent need for a quality evaluation method to assess transformer image quality, particularly to enhance the accuracy of transformer oil leakage identification tasks. Addressing this issue, this paper introduces image quality evaluation to the field of transformers for the first time and proposes a novel model called Multi-Directional Feature Extraction Transformer Image Quality Assessment (MFE-TIQA) for automatic evaluation of large-scale transformer images. MFE-TIQA comprises a main branch and a sub-branch. The main branch utilizes convolutional neural networks to extract multi-scale features from transformer images and employs a multi-directional feature extraction module to capture fine details. Meanwhile, the sub-branch employs a super-pixel segmentation model to generate local visual information from transformer images. Subsequently, a multi-branch information fusion module is constructed to comprehensively integrate the information from both branches, enhancing the model’s focus on critical details. Furthermore, comparative experiments are conducted on a self-constructed transformer image dataset and a publicly available dataset to validate the effectiveness of the proposed model. This research provides foundational evidence for subsequent transformer oil leakage identification and detection efforts.