Sample-specific repetitive learning for photo aesthetic auto-assessment and highlight elements analysis
- PDF / 2,072,936 Bytes
- 16 Pages / 439.37 x 666.142 pts Page_size
- 15 Downloads / 139 Views
Sample-specific repetitive learning for photo aesthetic auto-assessment and highlight elements analysis Ying Dai 1 Received: 18 September 2019 / Revised: 10 June 2020 / Accepted: 21 July 2020 # Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract
Aesthetic assessment is subjective, and the distribution of the aesthetic grades is overconcentrated in the middle levels. In order to realize the auto-assessment of photo aesthetics, we focus on using repetitive self-revised learning (RSRL) to retrain the convolutional neural network (CNN)-based aesthetics prediction network repetitively by the transfer learning, so as to improve the performance of imbalanced learning caused by the overconcentration distribution of aesthetic scores utilized as learning data. As RSRL, the network is trained repetitively by dropping out the low likelihood photo samples at the middle levels of aesthetics from the training data set based on the previously trained network. Further, the two retained networks are used in extracting aesthetic highlight elements of the photos to analyze the relation of the photo composition with the aesthetic assessment. The objective and subjective experimental results show that the CNN-based RSRL is effective for improving the performances of the imbalanced scores prediction network for the photos aesthetic auto-assessment. Keywords Photo aesthetic auto-assessment . Imbalanced learning . Repetitive self-revised learning . Dropping out sample . CNN
1 Introduction In response to the growth of digital camera, more and more pictures are taken to upload the social media. Many people hope to improve aesthetic level of themselves by taking beautiful photographs. So, auto-assessment of photo aesthetics is challenging. Researches have been investigating methods for providing automated aesthetical evaluation and classification of photographs. Aesthetic assessment is subjective. One of the main difficulties in addressing this challenge is in developing formal models of human aesthetic preference [1]. In this paper, authors stated that such models would allow computer systems to predict the aesthetic taste of * Ying Dai dai@iwate–pu.ac.jp
1
Iwate prefectural university, Takizawa, Japan
Multimedia Tools and Applications
a human being or adapt to the aesthetic tendencies of a human group. For making aesthetics automatic evaluation and choices, the best way to proceed is to create datasets for training the model in collaboration psychology aesthetics (PA) researchers, because computational aesthetics (CA) research typically reposts results using a success rate, while psychologists are more likely to use correlation. Closer collaboration between CA and PA can give rise to results that advance both disciplines. In [11], recent computer vision techniques used in the assessment of image aesthetic quality were reviewed. The authors discussed the possibility of manipulating the aesthetics of images through computational approaches. The research reviewed in the paper generally aims at assessing the aest
Data Loading...
