Título: Deep Learning-based Compression Artifacts Reduction for JPEG Image Classification
Autores: Andrey O. O. dos Reis, Edson M. Hung, Daniel G. Silva
Resumo: Computer vision has been one of the main application areas for Deep Learning (DL) techniques. Image classification, in particular, has been widely culminating in systems increasingly capable of replacing human visual analysis. However, having high-quality images is one of the main requirements for high performance of DL-based classification, but often factors associated with the models application condition make this ideal scenario impossible. Lossy compression fits very well in these conditions because it is a type of coding technique widely used in image storage and transmission. In this paper, we propose compression artifact reduction (CAR) as a way to circumvent the degradation problem for lossy compression. To this end, we perform experiments using JPEG with very low levels of quality factor (QF) compressing the test dataset for classification. Then, using a DL-based CAR model, we restore this same dataset in order to investigate a possible improvement in the classifiers performance. The two evaluated datasets presented positive results: in Flowers-102 we reached an average 0.52% increase in accuracy for 10 QFs values, whereas for Cub-200 this value was more expressive, around 34.57%. Those findings reinforce that DL-based CAR may increase performance in classification models degraded by drastic levels of signal corruption
Palavras-chave: deep learning, image classification, image restoration, compression artifacts, JPEG
Páginas: 7
Código DOI: 10.21528/CBIC2023-059
Artigo em pdf: CBIC_2023_paper059.pdf
Arquivo BibTeX: CBIC_2023_059.bib