研究目的
To enhance CNN-based no-reference image quality assessment by avoiding unreliable homogenous patches and biasing towards patches with complex structures using variance-based weighting.
研究成果
The proposed patch variance biased CNN approach significantly improves NR-IQA performance by avoiding homogenous patches and weighting towards high-variance patches, achieving state-of-the-art results on LIVE and TID2013 databases. This discovery opens new directions for further enhancements in CNN-based IQA methods.
研究不足
The method relies on patch variance as a homogenous indicator, which may not capture all aspects of image quality; computational time increases with higher variance thresholds; performance may vary with different types of distortions not fully covered in the databases.
1：Experimental Design and Method Selection:
The study uses a CNN architecture similar to Bosse's CNN with 12 convolutional layers, ReLU activation, max-pooling, dropout regularization, and ADAM optimizer with MAE loss. It involves patch sampling based on variance threshold, adaptive stride scanning for test patches, and variance-based weighted averaging for quality score estimation.
2：Sample Selection and Data Sources:
The LIVE database and TID2013 database are used, with 80% of images for training and 20% for testing. Patches are sampled from these images.
3：List of Experimental Equipment and Materials:
Computational setup using Keras with TensorFlow backend; no specific hardware mentioned.
4：Experimental Procedures and Operational Workflow:
Patches are sampled with a variance threshold (Tvar) during training; test patches are generated using adaptive stride scanning; quality scores are computed using weighted averaging based on patch variances.
5：Data Analysis Methods:
Performance is evaluated using Linear Correlation Coefficient (LCC) and Spearman Rank Order Correlation Coefficient (SROCC).
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
