Proceedings of 9th International Symposium on Graphic Engineering and Design最新文献

筛选

英文中文

CONTENT-AWARE IMAGE COMPRESSION WITH CONVOLUTIONAL NEURAL NETWORKS 基于卷积神经网络的内容感知图像压缩

Proceedings of 9th International Symposium on Graphic Engineering and Design Pub Date : 2018-11-01 DOI: 10.24867/GRID-2018-P56

Alen Selimović, A. Hladnik

{"title":"CONTENT-AWARE IMAGE COMPRESSION WITH CONVOLUTIONAL NEURAL NETWORKS","authors":"Alen Selimović, A. Hladnik","doi":"10.24867/GRID-2018-P56","DOIUrl":"https://doi.org/10.24867/GRID-2018-P56","url":null,"abstract":"Traditional image compression algorithms treat all image regions equally, regardless of their content, often resulting in reconstructed images that do not correlate well with human perception. Content-aware compression, on the other hand, prioritizes image regions that are more relevant to the interpretation of an image and encodes them at a higher bitrate, i.e. without loss or with less loss, than the rest of the image. Our paper explores the multi-structure region of interest (MS-ROI) model, a convolutional neural network, which enables the localization of multiple regions of interest (ROIs) in an image. The localization is expressed as a corresponding saliency map, which identifies the relevance of individual image regions and provides a saliency value for each pixel of the given image. This information is then used to guide the compression. The saliency values are discretized into multiple levels and more important levels are encoded with a higher quality factor Q than the less important ones, allowing for most of the reduction in image resolution to occur in non-salient image regions. Because the generated saliency maps produce soft boundaries between salient and non-salient image regions, smooth transitions between these regions are achieved. The obtained image is then encoded further using the standard JPEG algorithm with a uniform Q factor, resulting in the final image of the standard JPEG format. Our model was trained on the Caltech-101 image dataset and its performance was tested on two other image datasets. Presented are the obtained saliency maps for several images, as well as the results of contentaware compression, which are compared to the standard JPEG compression at different Q factors. For an objective comparison and evaluation of the quality of the obtained images, various standard quality metrics were used, i.e. mean squared error (MSE), peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) and multi-scale structural similarity index (MS-SSIM).","PeriodicalId":371126,"journal":{"name":"Proceedings of 9th International Symposium on Graphic Engineering and Design","volume":"72 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121191811","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 3

COMPRESSION RESISTANCE OF SMALL PAPERBOARD PACKAGING SHAPES 小纸板包装形状的抗压性

Proceedings of 9th International Symposium on Graphic Engineering and Design Pub Date : 2018-11-01 DOI: 10.24867/GRID-2018-P29

Josip Bota, Sonja Jamnicki Hanzer, D. Banić, M. Brozović

引用次数: 1

首页上一页