A Comparative Study of Feature Impact on a Perceptual Quality Evaluation in Smoky Laparoscopic Videos Using Artificial Neural Networks

Evaluating the impact of statistical features on a quality assessment of smoke-distorted laparoscopic videos is crucial as these distortions significantly impair evaluation accuracy. This work proposes a laparoscopic video quality assessment model that analyzes the effectiveness of luminance, the gray-level co-occurrence matrix (GLCM), and motion features in assessing video quality. The spectrally decomposed coefficients of these features and their interdependencies are modeled using univariate generalized Gaussian distribution (GGD) and bivariate GGD. An artificial neural network, trained with (a, b) GGD parameters and subjective scores, predicts frame-level quality, which is then averaged to obtain video-level quality. The results show that luminance features are significant in modeling video quality independently, while motion features perform better when combined with others. Considering feature interdependencies, especially between luminance and the GLCM, significantly improves the correlation between predicted and subjective scores. The proposed method, incorporating feature interdependencies, outperforms state-of-the-art approaches, underscoring its importance in replicating human perception.