Ischemic brain stroke occurs when a thrombus blocks a brain artery leading to a regional damage of brain due to lack of normal blood flow. Identification and diagnosis of stroke requires quick processing of medical image such as MRI. Segmentation of brain strokes is known to be a challenging task due to variability in the size and location of the stroke. In this paper, we examine a Transformer-based deep learning architecture, called Dual Attention-guided efficient Transformer (DAE-Former) for several stroke benchmarks. For efficient evaluation, several loss function combinations were examined along with dipping the original architecture, resulting in the Enhanced DAE-Former. The proposed modification leads to marginal improvement in the segmentation accuracy, especially evident in the results obtained from the ATLAS v2.0 dataset. In contrast, the performance improvements were not as pronounced with the smaller ISLES22 dataset. Both the original and Enhanced DAE-Former architectures performed similarly on this dataset. This outcome suggests that the enhanced model are more apparent in larger datasets, where the increased complexity of the architecture can be fully leveraged.
Cell-Free massive multiple-input multiple-output (MIMO) is gaining significant attention as a key area of interest in the beyond fifth-generation (B5G) and six-generation (6G) due to its ability to achieve better performance than traditional cellular systems. Cell-Free massive MIMO involves access points (APs) within a coverage area being connected to a single central processing unit (CPU) via fronthaul links, and these APs cooperated to support users. However, this system demands a significant amount of fronthaul capacity. To satisfy the finite fronthaul capacity, a system is configured using low-resolution ADCs. The consideration is for a user-centric Cell-Free massive MIMO system, which can reduce the fronthaul load compared to the traditional Cell-Free approach where all APs cooperate to support all users. Most user-centric cell-free massive MIMO research assumes a scenario where the number of APs is much greater than the number of users. In scenarios with a large number of users, a proposed clustering method ensures the connection between APs and users. Each user initially connects to the AP with the largest Large-Scale Fading Coefficient (LSFC). Subsequently, at each AP, a comparison of LSFC among users is performed, and the AP connects to the user with the largest LSFC. The proposed algorithm has confirmed an increase in both the average user spectral efficiency and the sum spectral efficiency when compared to the existing algorithm.
Quiz sheets are commonly used to measure students' learning outcomes. These sheets typically include test answers and grading marks made during review sessions. To make it easier to review and practice the questions again, it is suggested that these elements be removed and only the questions be kept on the quiz sheet. In this work, we propose using the Text Separation with Binarization Feature model (TSBFM) to identify the positions of the printed questions on the sheet. During the evaluation of model performance, we conducted tests on the self-collected dataset NtutQuizSheet. Our solution has shown an improvement of 7.81% in IoU metrics when compared to previous methods.
In the realm of digital media, the subtle interplay of colors in video content holds the key to unlocking viewer emotions. Our innovative study delves into this fascinating domain, leveraging the rich tapestry of colors in YouTube videos to decipher the emotional undertones they convey. By constructing a robust Color-Emotion Baseline Dictionary, we mapped specific colors to a spectrum of human emotions, uncovering the intricate ways in which visual hues influence viewer sentiment. Our methodology involved generating nuanced tints and shades, transcending conventional color-emotion models, and employing sophisticated barcode generation techniques for YouTube videos. This approach allowed for an unprecedented analysis of color psychology in digital media. In an innovative experiment using the Trailers12k dataset, we demonstrated the efficacy of our model, revealing significant correlations between color patterns and emotional responses. Our findings not only validate the profound impact of colors on emotional perception but also pave the way for novel applications in areas like digital marketing and content creation. Our work stands out as a unique contribution to the field, offering a fresh perspective on the emotional power of colors in multimedia content and setting a new benchmark for future research in this intriguing area.
To meet the demand of video transmission, learned video compression has drawn growing attention from research communities. In most existing works, motion estimation is performed through an optical flow prediction network. However, flow estimation naturally fails in complex cases such as occlusions, bringing in unexpected noise on warped temporal features. In order to tackle this issue, our work proposes a mask-based temporal context enhancement method. By explicitly analyzing the correlation of warped frame and original frame, a mask that reflects the credibility of flow warping is generated at encoder side and afterwards transmitted along with the motion information. By applying the mask on temporal features, we filter the noise and generate multi-scale temporal contexts with clearer semantics. Experimental results demonstrate that our model enjoys rate-distortion performance gains in terms of both PSNR and MS-SSIM over baseline contextual coding model.
