Panoramic Image Quality Assessment Based on Deep Learning

This article proposes a panoramic image quality evaluation model based on multi-layer perceptron. The model first constructs primary feature vectors that conform to human visual characteristics through color space decomposition and local contrast extraction, which serve as inputs for MLP. The core of the model lies in utilizing the deep nonlinear mapping capability of MLP to learn layer by layer from low-level distortion features to high-level quality semantic complex associations, in order to fit the highly nonlinear relationship between quality features and subjective ratings. To enhance perceptual alignment, the model converts the perceptible distortion characteristics into adaptive visual perception weighting factors, dynamically modulates the feature weights within the MLP, and achieves perceptual correction of uneven brightness areas. The experimental results show that the model significantly outperforms mainstream algorithms in terms of the correlation between quality prediction and subjective evaluation, verifying its superiority and robustness in objective quality assessment of panoramic images.