A No-Reference Image Quality Assessment Method Based on Random Forest for Power Scenario

No-reference image quality assessment (NR-IQA) is a crucial task in computer vision, aiming to evaluate the visual quality of images solely through distorted images without the need for original (reference) images. This evaluation method is particularly useful in practical applications where obtaining the original image is not feasible, such as in the task scenario discussed in this article: assessing whether a camera is blurred, overexposed, etc., in power grid scenarios. Traditional no-reference image quality assessment methods usually face issues such as over-reliance on manually designed features, which may not comprehensively capture all aspects of image quality, and good performance on specific datasets but significant performance degradation across different datasets or types of image distortion, resulting in poor consistency with actual subjective scores. Although with the continuous development of deep learning, many models are able to extract the features of an image with relatively high efficiency for effective evaluation, many of the models require high computational resources, which is not conducive to the needs of real-world scenarios where low-computing power is deployed to achieve timely image quality assessment. In order to solve these problems, we propose a random forest-based approach for reference-free image quality assessment in power scenarios. First, we collect original power scene images and perform image expansion, including abnormal and normal images, to increase the diversity of the dataset. The expanded images were merged with the original images to form a balanced dataset. Next, we take the merged dataset as input and perform data processing, including random cropping and simulated image damage, to enhance the model's adaptability to different observation viewpoints and damage scenarios. The processed images are used for feature extraction, and these features are subsequently used to train the random forest model. The experimental results show that our method performs well in this scenario and has low computational power requirements, which is conducive to a wide range of applications in future power scenarios and provides an effective technological tool for university monitoring in the power industry.