Intelligent credibility evaluation calculation method for visual models for electric equipment inspection

During the power inspection process, image and video data are generated, and building visual models based on these data for defect recognition and fault diagnosis has been widely applied in the power industry. Due to external factors such as weather and transportation, adversarial samples have become an important factor in reducing the efficiency of visual models. In response to this issue, a credibility evaluation method for power vision models is proposed in this paper. Analyze the mechanism of adversarial sample generation in image and video data generated during power inspection and extract common features of adversarial samples through cross sample comparison and matching. Then, based on the accuracy and efficiency requirements of visual models for power application scenarios such as transmission channel monitoring, drone inspection, and safety monitoring, design visual model efficiency and credibility evaluation rules. Building a visual model credibility evaluation model based on the efficiency and credibility evaluation rules of electric power visual models, to calculate the accuracy, precision, recall of visual object classification tasks, as well as the average accuracy and recall of detection tasks. Calculate the interference intensity of adversarial samples based on the weight calculation results, form the corresponding credibility evaluation report of the power vision model, and feedback it to the data management party, thereby achieving closed-loop management of the basic data quality of the power vision model, improving the sample classification effect and recognition accuracy of the power vision model.