基于多尺度频率注意力融合的端子文本检测方法OA

The detection of terminal numbers is a very critical task in the acceptance process of substation secondary system cables.A terminal text detection method that incorporates a multi-scale frequency attention mechanism is proposed,which aims to improve the efficiency and accuracy of number checking.DBNet is used as the base model,and improvements are made to address the model's shortcomings in feature extraction and fusion in complex scenes.Terminal images in real-world situations often suffer from blurriness,low resolution,and strong background interference,which poses a challenge to traditional detection algorithms,especially in terms of their limited ability to extract high-quality features.In order to solve this problem,the module AFT-FFE is designed,which combines the adaptive Fourier transform with the frequency feature enhancement technique.This module enables the model to capture the high-frequency details and structural features in the image more effectively by introducing frequency domain analysis,which improves the model's ability to localize the terminal text.In the second aspect,the traditional feature fusion process is prone to the problems of information loss and scale confusion,and the MSF-KAN module is introduced.The module incorporates a nonlinear attention mechanism and a dynamic weighting strategy,which can achieve adaptive fusion and emphasize key information among multi-scale features,and helps to improve the performance stability and detection accuracy of the model in complex backgrounds.For the experimental part,validation is performed on the private dataset ZET.The results show that the proposed method outperforms multiple mainstream algorithms in terms of precision,recall and HMean value,reaching 91.22%,88.52%and 89.85%,respectively.Results verifiy that the method has the good robustness and practical value of the proposed method in practical applications,which is expected to provide strong technical support for scenarios such as wiring inspection of power equipment.