Zhenchang Zhang

Accurate prediction of Sea Surface Temperature (SST) plays a critical role in climate research and marine ecosystem management. Traditional models predict trends by analyzing and fitting data, but they struggle with capturing long-range dependencies and complex spatiotemporal patterns. The Transformer’s attention mechanism effectively addresses long-range dependencies, but its high computational complexity poses challenges. To overcome these limitations, this study proposes a novel spatiotemporal sequence prediction model: the Spatiotemporal Vision Mamba Recurrent Neural Network (SVRNN). The model innovatively integrates a bidirectional state-space processing mechanism and decoupled memory modules. The bidirectional mechanism maintains a global receptive field with linear computational complexity, while the decoupled memory modules explicitly separate spatiotemporal dependencies, enhancing the model’s ability to capture complex spatiotemporal patterns. During the experiment on hourly SST prediction in the Taiwan Strait, where the SST of the next 12 hours was predicted using data from the previous 12 hours, the SVRNN model demonstrated superior performance, achieving a Root Mean Square Error (RMSE) of 0.159°C, a Mean Absolute Error (MAE) of 0.105°C, and a Mean Absolute Percentage Error (MAPE) of 0.496%. Furthermore, our seasonal error analysis reveals that the model exhibits robust performance in different seasons, providing more reliable technical support for SST prediction in the Taiwan Strait.

Study the correlation between the multi-scale entropy values of electroencephalo-gram (EEG) signals in medical personnel and the degree of mental fatigue, and analyze the detection of mental fatigue status in healthcare workers. Utilizing singular value analysis to eliminate noise in EEG signal data, calculating multi-scale entropy values of the original EEG signals, referencing the FS-4 fatigue scale, and analyzing entropy value stability; proposing three fatigue analysis algorithms?random forest, BP neural network, and XGBoost comparing them with traditional machine learning methods and artificial neural network methods , investigating the relationship between multi-scale entropy values of EEG 1 signals and fatigue in healthcare personnel, and analyzing the determination of mental fatigue status. This paper employs the multi-scale entropy method to extract multi-scale information from original EEG signals. The combination of multi-scale information with three fatigue analysis algorithms enables the detection and analysis of mental fatigue status in healthcare personnel. Experimental results indicate that the XGBoost algorithm exhibits the highest sensitivity at 75%, while the random forest algorithm demonstrates the best specificity at 80% and an AUC value of 95%. The random forest algorithm demonstrates superior performance in studying mental fatigue.

In conventional image super-resolution tasks, the primary elements targeted for enhancement are objects in natural scenes, which convey predominantly visual texture information. However, text in scene images differs from these objects in containing two levels of content: visual texture and semantic information. Unfortunately, most natural scene text super-resolution algorithms treat text as a regular natural object without considering its semantic information. This paper proposes a Multimodal Super-Resolution Neural Network (MSRNN) combining text and image information. Specifically, we extract semantic information from text images using text knowledge and utilize this to guide the super-resolution network in enhancing low-resolution text images. Additionally, we employ a visual graph convolutional network (VIG) to extract image features while preserving structural information. Furthermore, we design a contextual orthogonal attention module to effectively extract the structural and edge information of text images and integrate the text-image features for recognition. Experimental results demonstrate that our model achieves state-of-the-art performance among existing models.

Knowledge distillation (KD) is a well-established technique for compressing neural networks and has gained increasing attention in object detection tasks. However, typical object detection distillation methods use fixed-level semantic features for distillation, which might not be best for all training stages and samples. In this paper, a multilayer semantic feature adaptive distillation (MSFAD) method is proposed that uses a routing network composed of a teacher and a student detector, along with an agent network for decision making. Specifically, the inputs to the proxy network consist of the features output by the neck structures of the teacher and student detectors, and the output is a decision on which features to choose for distillation. The MSFAD method improves the distillation training process by enabling the student detector to automatically select valuable semantic-level features from the teacher detector. Experimental results demonstrated that the proposed method increased the mAP₅₀ of YOLOv5s by 3.4% and the mAP_50–90 by 3.3%. Additionally, YOLOv5n with only 1.9 M parameters achieved detection performance comparable to that of YOLOv5s.

Semi-supervised learning is a learning pattern that can utilize labeled data and unlabeled data to train deep neural networks. In semi-supervised learning methods, self-training-based methods do not depend on a data augmentation strategy and have better generalization ability. However, their performance is limited by the accuracy of predicted pseudo-labels. In this paper, we propose to reduce the noise in the pseudo-labels from two aspects: the accuracy of predictions and the confidence of the predictions. For the first aspect, we propose a similarity graph structure learning (SGSL) model that considers the correlation between unlabeled and labeled samples, which facilitates the learning of more discriminative features and, thus, obtains more accurate predictions. For the second aspect, we propose an uncertainty-based graph convolutional network (UGCN), which can aggregate similar features based on the learned graph structure in the training phase, making the features more discriminative. It can also output the uncertainty of predictions in the pseudo-label generation phase, generating pseudo-labels only for unlabeled samples with low uncertainty; thus, reducing the noise in the pseudo-labels. Further, a positive and negative self-training framework is proposed, which combines the proposed SGSL model and UGCN into the self-training framework for end-to-end training. In addition, in order to introduce more supervised signals in the self-training process, negative pseudo-labels are generated for unlabeled samples with low prediction confidence, and then the positive and negative pseudo-labeled samples are trained together with a small number of labeled samples to improve the performance of semi-supervised learning. The code is available upon request.