研究目的
To propose a novel image quality assessment method based on variable receptive field and information entropy to better measure multi-scale objects in images, addressing the limitations of traditional methods that ignore variable spatial configuration information.
研究成果
RSEI effectively measures image quality by utilizing semantic segmentation and information entropy, outperforming state-of-the-art methods in consistency with subjective evaluations, and is suitable for scenarios like satellite images where ground truth is unavailable.
研究不足
The method increases running time due to semantic segmentation and depends on the accuracy requirement for the number of image patches, which may not be optimal without mean opinion scores in real scenarios. Deep learning-based methods require large datasets and hardware acceleration.
1：Experimental Design and Method Selection:
The method involves semantic segmentation of images using superpixel segmentation to divide images into patches, calculation of weights based on information entropy, and use of mutual information to assess quality.
2：Sample Selection and Data Sources:
The TID2008 database with 25 reference images and 1700 distorted images, and the FEI faces database with 400 images, are used.
3：List of Experimental Equipment and Materials:
Computer with Intel Core i5-6300HQ CPU, 8 GB RAM, NVIDIA GeForce GTX 950M GPU.
4：Experimental Procedures and Operational Workflow:
Images are converted to YCbCr color space, segmented into patches using superpixel segmentation, normalized to rectangles, weights are calculated, mutual information is computed, and RSEI is derived.
5：Data Analysis Methods:
Performance is evaluated using Kendall rank-order correlation coefficient (KROCC), Spearman rank-order correlation coefficient (SROCC), Pearson linear correlation coefficient (PLCC), and root mean-squared error (RMSE).
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
