CAAI Transactions on Intelligence Technology最新文献

Existing weakly supervised semantic segmentation (WSSS) methods based on image-level labels always rely on class activation maps (CAMs), which measure the relationships between features and classifiers. However, CAMs only focus on the most discriminative regions of images, resulting in their poor coverage performance. We attribute this to the deficiency in the recognition ability of a single classifier and the negative impacts caused by magnitudes during the CAMs normalisation process. To address the aforementioned issues, we propose to construct selective multiple classifiers (SMC). During the training process, we extract multiple prototypes for each class and store them in the corresponding memory bank. These prototypes are divided into foreground and background prototypes, with the former used to identify foreground objects and the latter aimed at preventing the false activation of background pixels. As for the inference stage, multiple prototypes are adaptively selected from the memory bank for each image as SMC. Subsequently, CAMs are generated by measuring the angle between SMC and features. We enhance the recognition ability of classifiers by adaptively constructing multiple classifiers for each image, while only relying on angle measurement to generate CAMs can alleviate the suppression phenomenon caused by magnitudes. Furthermore, SMC can be integrated into other WSSS approaches to help generate better CAMs. Extensive experiments conducted on standard WSSS benchmarks such as PASCAL VOC 2012 and MS COCO 2014 demonstrate the superiority of our proposed method.

Recently, learning-based control for multi-robot systems (MRS) with obstacle avoidance has received increasing attention. The goals of formation control and obstacle avoidance could be intrinsically tied. As a result, developing a safe and near-optimal control policy with the actor-critic structure is challenging. Therefore, a hybrid distributed and decentralised asynchronous actor-critic reinforcement learning (Di-De-RL) technique is proposed to address this problem. First, we decompose the integrated formation control and collision avoidance problem into two successive ones. To solve them, we design a distributed reinforcement learning (Di-RL) algorithm that employs a neural network-based actor-critic structure for formation control, and a decentralised RL (De-RL) algorithm that incorporates a potential-field (PF)-based actor-critic structure for collision avoidance. In Di-RL, the actor-critic pairs are trained in a distributed manner to achieve near-optimal consensus formation control. With the trained policy of Di-RL fixed, the PF actor-critic pairs in De-RL are trained in a decentralised manner for safe collision avoidance. Such an asynchronous training design of the hybrid Di-RL and De-RL enables weight convergence and control safety in the learning process. The simulated and real-world experimental results demonstrate the effectiveness and enhanced performance of the approach in formation control with both static and dynamic obstacle avoidance, highlighting its advantages in resolving the conflict between the safety objective and optimal control.

Functional brain networks have been used to diagnose brain disorders such as autism spectrum disorder (ASD) and attention-deficit/hyperactivity disorder (ADHD). However, existing methods not only fail to fully consider various levels of interaction information between brain regions, but also limit the transmission of information among unconnected regions, resulting in the node information loss and bias. To address these issues, we propose a causality-guided multi-view diffusion (CG-MVD) network, which can more comprehensively capture node information that is difficult to observe when aggregating direct neighbours alone. Specifically, our approach designs multi-view brain graphs and multi-hop causality graphs to represent multi-level node interactions and guide the diffusion of interaction information. Building on this, a multi-view diffusion graph attention module is put forward to learn node multi-dimensional embedding features by broadening the interaction range and extending the receptive field. Additionally, we propose a bilinear adaptive fusion module to generate and fuse connectivity-based features, addressing the challenge of high-dimensional node-level features and integrating richer feature information to enhance classification. Experimental results on the ADHD-200 and ABIDE-I datasets demonstrate the effectiveness of the CG-MVD network, achieving average accuracies of 79.47% and 80.90%, respectively, and surpassing state-of-the-art methods.

Obesity is a major risk factor for chronic diseases, underscoring the need for early diagnosis and effective management. This study presents a novel expert system designed to accurately classify obesity levels and provide personalised treatment recommendations. Five machine learning algorithms—decision tree, random forest, multinomial logistic regression (MLR), Naive Bayes, and support vector machine (SVM)—were evaluated using the SEMMA data mining methodology and the tidymodels framework. MLR demonstrated the highest accuracy (97.48%) and was selected as the final model. The system features a user-friendly interface built with R Shiny, facilitating real-time interaction and a seamless user experience. Treatment recommendations are generated through if-then rule-based logic, ensuring tailored guidance for each obesity category. Comparative analysis highlights the system's superior diagnostic accuracy and practical application in treatment guidance. Its accessibility, particularly in underserved rural populations, enhances public health outcomes by enabling early diagnosis, targeted interventions, and proactive obesity management.

Robotics plays an increasingly important role in all areas of human activity. Teleoperation robots can effectively ensure the safety of operators when operating in difficult and high-risk industrial scenarios, which obviously requires instant and efficient signal compression and transmission in the system. However, most of the existing algorithms cannot fully explore the correlation within the signal, which mostly limits the compression efficiency. In this paper, a novel prediction-aided kinaesthetic-signal compression framework is proposed, which uses semantic communication methods to explore the temporal and spatial correlations of signals and employs neural network predictions to uncover their internal correlations. Specifically, the signal is first divided into two groups: the base part and the predictable part, and then a series of transformation matrices are introduced to establish the correlation between the two groups of the signal, which can be automatically optimised by a well-designed neural network. This strategy of using learnable transformation matrices for prediction can not only accurately construct the correlation within the signal through massive data mining but also efficiently execute inference in a simple matrix multiplication computing form. Experimental results demonstrate that the proposed method outperforms the existing traditional tactile codecs and the latest tactile semantic communication methods.

The evaluation and assessment of network security is a decision-making (DM) problem that occurs in an environment with multiple criteria, which have uncertainty, bipolarity, and extra-related information. The traditional approaches fail to address the need to acquire a wide range of information for the assessment, especially in situations where the criteria have both positive and negative aspects and contain extra fuzzy information. Therefore, in this manuscript, we aim to introduce a DM approach based on the concept of bipolar complex fuzzy (BCF) Yager aggregation operators (AOs). The related properties of these aggregation operators (AOs) are also discussed. Moreover, in this article, we diagnose the Yager operations in the setting of BCF. The basic idea of the interpreted operators and DM approach is to access the problem linked with the network security that is to evaluate and select the finest network security control and network security protocols for protecting and safeguarding the network of any organization or home (case studies). Finally, to exhibit the supremacy and success of the described theory, we examine them with the prevailing theories.

Underwater acoustic target recognition (UATR) has become increasingly prevalent for ocean detection, localisation, and identification. However, due to the complexity and variability of underwater environments, especially in multi ship event environments, where multiple acoustic signals coexist, practical applications face significant challenges. These challenges hinder single-category acoustic recognition algorithms, particularly in extracting time series features and achieving fine-grained or multi-scale feature fusion. This paper innovatively introduce the SKANN framework, which achieve precise submarine sound recognition in underwater mixed ship events environments through timing data enhancement and sampling training module and selective kernel feature extraction module. The timing data enhancement and sampling training module improves time sequence feature extraction through progressive acoustic sampling. The selective kernel feature extraction module effectively fuses multi-scale features by integrating selective kernel (SK) technology. To simulate concurrent ship events, we constructed the mixed ship noise dataset (M-DeepShip), providing an experimental basis and test platform for underwater mixed ship event detection. This dataset ensures that the model encounters diverse audio samples during training and validation, improving its ability to extract temporal features. Experimental results show that SKANN achieves a 93.6% recognition rate on the M-DeepShip dataset, demonstrating its effectiveness in recognising underwater mixed ship events. Given the complexity of real underwater environments, this work lays a crucial foundation for the sound recognition of submarine vessels. Future research will focus on real marine environments to validate and refine the models and methods for practical applications.

Few-shot learning is the task of identifying new text categories from a limited set of training examples. The two key challenges in few-shot learning are insufficient understanding of new samples and imperfect modelling. The uniqueness of low-resource languages lies in their limited linguistic resources, which directly leads to the difficulty for models to learn sufficiently rich feature representations from limited samples. As a minority language, Tibetan few-shot learning requires further exploration. With limited data resources, if the model's understanding of text is noncontextual, it cannot provide sufficiently distinctive feature representations, limiting its performance in few-shot learning. Therefore, this paper proposed a few-shot learning architecture called two-level word embeddings matching networks (TWE-MN). TWE-MN is specifically designed to enhance the model's representational capacity and optimise its generalisation capabilities in data-scarce environments. As this paper focuses on Tibetan few-shot learning tasks, a pretrained Tibetan language model, BoBERT, was constructed. BoBERT, as the pre-embedding layer of TWE-MN, in combination with the BoBERT-augmented full-context embedding, can capture feature information from local to global levels. This paper evaluated the performance of TWE-MN in Tibetan few-shot learning tasks and Tibetan text classification tasks. The experimental results show that TWE-MN outperformed vanilla MN in all Tibetan few-shot learning tasks, with an average accuracy improvement of 4.5%–6.5% and up to 6.8% at most. In addition, this paper also explores the potential of TWE-MN in other NLP tasks, such as text classification and machine translation.

Segmentation tasks require multiple annotation work which is time-consuming and labour-intensive. How to make full use of unlabelled data to assist in training deep learning models has been a research hotspot in recent years. This paper takes instrument segmentation in endoscopic surgery as the background to explore how to use unlabelled data for semi-supervised learning more reasonably and effectively. An adaptive gradient correction method based on the degree of perturbation is proposed to improve segmentation accuracy. This paper integrates the recently popular segment anything model (SAM) with semi-supervised learning, taking full advantage of the large model to enhance the zero-shot ability of the model. Experimental results demonstrate the superior performance of the proposed segmentation strategy compared to traditional semi-supervised segmentation methods, achieving a 2.56% improvement in mean intersection over union (mIoU). The visual segmentation results show that incorporation of SAM significantly enhances our method, resulting in more accurate segmentation boundaries.