This paper presents a new technique to embed information in 3D printed objects using a near infrared fluorescent dye. Regions containing a small amount of fluorescent dye are formed inside an object during fabrication to embed information inside it, and these regions form a pattern that expresses certain information. When this object is irradiated with near-infrared rays, they pass through the object made of resin but are partly absorbed by the fluorescent dye in the pattern regions, and it emits near-infrared fluorescence. Therefore, the internal pattern can be captured as a high-contrast image using a near-infrared camera, and the embedded information can be nondestructively read out. This paper presents a technique of forming internal patterns at two different depths to double the amount of embedded information. We can determine the depth of the pattern from the image because the distribution of the brightness of the captured image of the pattern depends on its depth. We can also know from the brightness distribution whether or not a pattern exists at two depths. Because this can express four states, the amount of embedded information can be doubled using this method. We conducted experiments using deep learning to distinguish four states from the captured image. The experiments demonstrated the feasibility of this technique by showing that accurate embedded information can be read out.

In this study, to promote the translation and digitization of historical documents, we attempted to recognize Japanese classical ‘kuzushiji’ characters by using the dataset released by the Center for Open Data in the Humanities (CODH). ‘Kuzushiji’ were anomalously deformed and written in cursive style. As such, even experts would have difficulty recognizing these characters. Using deep learning, which has undergone remarkable development in the field of image classification, we analyzed how successfully deep learning could classify more than 1,000-class ‘kuzushiji’ characters through experiments. As a result of the analysis, we identified the causes of poor performance for specific characters: (1) ‘Hiragana’ and ‘katakana’ have a root ‘kanji’ called ‘jibo’ and that leads to various shapes for one character, and (2) shapes for hand-written characters also differ depending on the writer or the work. Based on this, we found that it is necessary to incorporate specialized knowledge in ‘kuzushiji’ in addition to the improvement of recognition technologies such as deep learning.

In Hematoxylin and Eosin (H&E) stained images, it is difficult to distinguish collagen and elastic fibers because these are similar in color and texture. This study tries to segment the appearance of elastic and collagen fibers based on U-net deep learning using spatial and spectral information of H&E stained hyperspectral images. Groundtruth of the segmentation is obtained using Verhoeff’s Van Gieson (EVG) stained images, which are commonly used for recognizing elastic and collagen fiber regions. Our model is evaluated by three cross-validations. The segmentation results show that the combination of spatial and spectral features in H&E stained hyperspectral images performed better segmentation than H&E stained in conventional RGB images compare to the segmentation of EVG stained images as ground truth by visually and quantitatively.

As a consequence of a lack of balance between the levels of difficulty of a game and the players’ skills, the resulting experience for players might be frustrating (too difficult) or boring (too easy). Players having a bad experience could impact game creators negatively, leading to irreparable damage. The main motivation of this study was to find effective ways to reduce the gap between skills and difficulty,to help developers create a more suitable experience for players. This paper shows the results of applying Neural Networks and Support Vector Machines to data collected from the pressure exerted to a gamepad’s button with the purpose of finding patterns that can help predict: difficulty, fun, frustration, boredom,valence, arousal and dominance at a determined time. We obtained results with an accuracy of 83.64 % when predicting boredom, around 70 % of accuracy classifying frustration, fun, difficulty and dominance.

In this paper, we are tackling the problem of the entire 360-degree image generation process by viewing one specific aspect of it. We believe that there are two key points for achieving this 360-degree image completion: firstly, understanding the scene context; for instance, if a road is in front, the road continues on the back side; and secondly, the treatment of the area-wise information bias; for instance, the upper and lower areas are sparse due to the distortion of equirectangular projection, and the center is dense since it is less affected by it. Although the context of the whole image can be understood through using dilated convolutions in a series, such as recent conditional Generative Adversarial Networks (cGAN)-based inpainting methods, these methods cannot simultaneously deal with detailed information. Therefore, we propose a novel generator network with multi-scale dilated convolutions for the area-wise information bias on one 360-degree image and a self-attention block for improving the texture quality. Several experiments show that the proposed generator can better capture the properties of a 360-degree image and that it has the effective architecture for 360-degree image completion.

K-SVD (K-Singular Value Decomposition) is a popular technique for learning a dictionary that offers sparse representation of the input data, and has been applied to several image coding applications. It is known that K-SVD performance is largely dependent on the features of the training images. Therefore, a multi-class dictionary approach is appropriate for natural images given the variety of their features. However,most published investigations of multi-class dictionaries are based on predetermined classification and do not consider the relation between classification stage and dictionary training stage. Therefore, there is still room for improving coding efficiency by linking dictionary training with classification optimization. In this paper,we propose a multi-class dictionary design method that repeats the following two stages: class update stage for all training vectors and dictionary update stage for each class by K-SVD. Experiments indicate that the proposed method outperforms the conventional alternatives as it achieves, for the fixed classification task,BD-bitrate scores of 6% to 48% and the BD-PSNR value of 0.4 dB to 1.6 dB.

This paper presents a fixed-complexity inter mode filtering algorithm for High Efficiency Video Coding (HEVC).Motion estimation (ME) particularly the fractional motion estimation (FME) is the most computational challenge part because of the optimized Rate-Distortion (RD) evaluation for coding blocks with different sizes and prediction modes. Especially when HEVC introduces new tools like the coding tree unit and asymmetric motion partition (AMP), complexity of ME has increased furthermore. There are many approaches to reducing the ME complexity by removing the unlikely prediction modes. However, most mode filtering methods cannot guarantee the worst-case performance and limit overall speed in a pipelined video encoder.Thus, a fixed number of modes is required to ensure the worst-case performance and realize the pipeline design. By investigating the cost correlation of integer motion estimation (IME) and FME, we utilize the confidence interval (CI) of the cost ratio to remove fixed number of modes. Moreover, the dedicated configurable mode filtering makes the complexity of FME adaptive to different requirements of RD performance. According to experiment in HEVC reference software HM 16.0 with full FME, the proposed scheme achieves at almost 82.75% complexity reduction of the FME with an average of 1.63% BD rate loss.