Transformer-Based No-Reference Image Quality Assessment via Supervised Contrastive Learning

Authors: Jinsong Shi, Pan Gao, Jie Qin

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on six standard IQA datasets show that Sa TQA outperforms the state-of-the-art methods for both synthetic and authentic datasets.
Researcher Affiliation Academia Jinsong Shi, Pan Gao*, Jie Qin* College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics
Pseudocode No The paper describes methods with figures and mathematical formulas but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/I2-Multimedia-Lab/Sa TQA.
Open Datasets Yes For self-supervised pre-training, we use the KADIS (Lin, Hosu, and Saupe 2020) dataset. ... The synthetic datasets include CSIQ (Larson and Chandler 2010), TID2013 (Ponomarenko et al. 2015), and KADID10K (Lin, Hosu, and Saupe 2019), and the authentic datasets include LIVE Challenge (Ghadiyaram and Bovik 2015), KONIQ (Hosu et al. 2020), and LIVE-FB (Ying et al. 2020).
Dataset Splits No The paper mentions training and testing but does not explicitly provide specific percentages or counts for training/validation/test dataset splits.
Hardware Specification Yes In this paper, all of our experiments are performed using Py Torch on a single NVIDIA Ge Force RTX3090.
Software Dependencies No The paper mentions 'Py Torch' but does not specify a version number or other software dependencies with version numbers.
Experiment Setup Yes During image preprocessing, we cropped 8 random 224x224 patches from each image and randomly flipped the cropped patches horizontally. ... In the training process, we use Adam W optimizer with learning rate of 2 10 5, weight decay of 1 10 2, and learning strategy of cosine annealing, where Tmax is set to 50 and etamin is 0. Experiments are trained for 150 epochs. The loss function used is L1Loss, and the batch size is 4.