Target speaker extraction (TSE) acts as a front-end processing technology for various speech applications, such as automatic speech recognition. However, TSE has long faced challenges from under-determined environments and in the presence of noise. In this paper, we propose a dual-channel system for directional TSE under noisy underdetermined conditions. In our approach, we utilize two source models that integrate conditional variational autoencoders (CVAEs) with global style tokens (GSTs) to learn representations of the noisy single speech and the noisy mixed speech within a geometric source separation framework, where GSTs generate conditional variables for CVAEs. To address residual noise in the extracted target signal under various noisy conditions, we introduce a conditional neural postfilter with a GST to estimate a complex Time-Frequency (T-F) mask for denoising. Additionally, we propose a joint network, where a conditional neural postfilter is jointly trained with a CVAE and a shared GST module. The experimental results demonstrate that our proposed dual-channel TSE method achieves better performance under noisy underdetermined conditions.

Molecular Representation Learning (MRL) is widely applied in various downstream tasks, such as molecule generation, molecular property prediction and reaction prediction. Nevertheless, MRL faces several challenges posed by the vast chemical space and limited labeled-data availability. In this paper, we propose Hierarchical Graph Transformer (HieGT), integrating atom-level and motif-level representations to capture local-global characteristics of molecules over a hierarchical graph. Leveraging 2D topological and 3D geometric encoding, HieGT enhances intrinsic representation understanding of molecules. The proposed method achieves the state-of-the-art performance over the molecular property prediction dataset PCBA of Open Graph Benchmark (OGB), and competitive results on PCQM4Mv2 with better interpretability.

The Moving Picture Experts Group (MPEG) has published the geometry-based point cloud compression (G-PCC) standard. It converts the compression of irregular point coordinates to the coding of structured binary octree node occupancy, where the Context-based Adaptive Binary Arithmetic Coding (CABAC) can be applied. The context model, constructed by intra and inter-octree layer information, drives the probability update of the arithmetic coder with a so-called Optimal Binarization with Update On-the-Fly (OBUF) scheme. The original OBUF design, while effective, lacks a probability range limitation for each binary coder, leading to issues in probability estimation accuracy and convergence speed. Moreover, when coding dynamic point clouds, the inter-frame information is not efficiently considered in OBUF, leading to excessive memory consumption for storing and tracking context states. To address these challenges, we propose an initialization strategy for both fine-grained context states (Fine-CtxS) and coarse-grained context states (Coarse-CtxS) in OBUF, alongside an adaptive probability bound determination method for each Coarse-CtxS to confine probability estimation. Furthermore, the paper delves into improvements for inter-frame geometry coding, including the construction of Fine-CtxS, and reducing memory consumption of Fine-CtxS in OBUF. The proposed methods have been adopted in recent G-PCC Edition 2 standardization activities, demonstrating enhanced performance.

We introduce GreenCOD, a green method for detecting camouflaged ob jects distinct in its avoidance of backpropagation techniques. GreenCOD leverages gradient boosting and deep features extracted from pre-trained Deep Neural Networks. Traditional camouflaged object detection approaches rely on complex deep neural networks, seeking performance improvements by backpropagation-based finetuning. However, such methods are typically computationally demanding and exhibit only marginal performance variations across different models. It raises the question of whether effective training can be achieved without backpropagation. In this direction, our work proposes a new paradigm that utilizes gradient boosting for COD. This approach significantly simplifies the model design, resulting in a system that requires fewer parameters and operations and maintains high performance compared to state-of-the-art deep learning models. Remarkably, our models are trained without backpropagation and achieve the best performance with fewer than 20G Multiply-Accumulate Operations. This new, more efficient paradigm opens avenues for further exploration in green, backpropagation-free model training. We make GreenCOD source code and on-device demo available at https://greencod.ai/ for futher research.

Face Super-Resolution (FSR) represents a significant branch of image super-resolution, aiming to reconstruct low-resolution face images into high-resolution counterparts. Recently, driven by rapid advancements in deep learning technology, FSR methods using deep learning have achieved notable subjective and objective reconstruction quality, attracting extensive industrial attention. However, detailed classifications of FSR methods remain limited. Therefore, this survey systematically and comprehensively reviews deep learning-based FSR methods. Initially, we introduce the background and technical framework of FSR. Subsequently, we detail the FSR problem definition, alongside commonly used datasets, evaluation metrics, and loss functions. We conduct comprehensive researches in deep learning FSR methods and classify them according to their solution strategies. Within each category, we begin with a general method description, and subsequently introduce representative approaches and discuss their respective pros and cons. Finally, we address current challenges in FSR methods and propose future research directions.

The number of people with mental illness is increasing because of stress or environmental influences. They require close monitoring because of their unpredictable behaviors; however, this is challenging given the lack of adequate numbers of medical staff. To overcome this problem, we propose a personalized abnormal behavior recognition alarm system for closed wards. The proposed system utilizes real-time video analysis to detect and track the locations of the patients, enabling recognition of their abnormal behaviors. In addition, new definitions are provided for specific abnormal behaviors that commonly occur in closed wards, with the adaptation of continual learning in the system. This architecture allows the creation of an abnormal behavior dataset while enhancing the recognition accuracy. The average abnormal behavior recognition accuracy with this system is over 92%. According to test results in real hospitals, about 84% of the medical staff were satisfied with the proposed system. Through the proposed alarm system, the staff could implement immediate actions without careful monitoring. By reducing the probability of occurrence of dangerous incidents, the system not only benefits the health of the patients but also enhances the working environment of the medical staff.