Method for Identifying Power Identification Board Under Tilt Angle Based on Geometric Correction and Attention Enhancement

Power equipment identification plates carry critical information such as equipment model and voltage level. Accurate and efficient recognition of these plates is crucial for power grid safety operation and maintenance. However, in practical inspections, identification plate images often suffer from issues such as tilting, rotation, multi-scale variations, and uneven lighting due to shooting angles and installation positions, severely impacting the performance of traditional recognition methods. This paper addresses these challenges by proposing a deep learning-based method for recognizing tilted power equipment identification plates. The method first uses multi-directional and multi-scale Log-Gabor filter banks to extract image key points and generate spatial mask information, employs a ResNet-based feature extraction method to obtain multi-scale features, and applies adaptive dynamic weighting to fuse features with spatial mask information, enhancing adaptability to small targets and complex scenes. Then, a differentiable binarization method is introduced to precisely segment the integrated feature map, obtaining a binary mask of the identification plate region. Based on this mask, the center of mass is calculated and polar coordinate transformation is performed, combined with Log-Gabor filter banks to generate rotation-invariant multi-scale Log-Gabor descriptors. Finally, the generated descriptors and center position are used to correct image tilt, and a character recognition method based on bidirectional recurrent neural network with.