CAAI Transactions on Intelligence Technology最新文献

This paper introduces a novel ensemble Deep learning (DL)-based Multi-Label Retinal Disease Classification (MLRDC) system, known for its high accuracy and efficiency. Utilising a stacking ensemble approach, and integrating DenseNet201, EfficientNetB4, EfficientNetB3 and EfficientNetV2S models, exceptional performance in retinal disease classification is achieved. The proposed MLRDC model, leveraging DL as the meta-model, outperforms individual base detectors, with DenseNet201 and EfficientNetV2S achieving an accuracy of 96.5%, precision of 98.6%, recall of 97.1%, and F1 score of 97.8%. Weighted multilabel classifiers in the ensemble exhibit an average accuracy of 90.6%, precision of 98.3%, recall of 91.2%, and F1 score of 94.6%, whereas unweighted models achieve an average accuracy of 90%, precision of 98.6%, recall of 93.1%, and F1 score of 95.7%. Employing Logistic Regression (LR) as the meta-model, the proposed MLRDC system achieves an accuracy of 93.5%, precision of 98.2%, recall of 93.9%, and F1 score of 96%, with a minimal loss of 0.029. These results highlight the superiority of the proposed model over benchmark state-of-the-art ensembles, emphasising its practical applicability in medical image classification.

Realistic human reconstruction embraces an extensive range of applications as depth sensors advance. However, current state-of-the-art methods with RGB-D input still suffer from artefacts, such as noisy surfaces, non-human shapes, and depth ambiguity, especially for the invisible parts. The authors observe the main issue is the lack of geometric semantics without using depth input priors fully. This paper focuses on improving the representation ability of implicit function, exploring an effective method to utilise depth-related semantics effectively and efficiently. The proposed geometry-enhanced implicit function enhances the geometric semantics with the extra voxel-aligned features from point clouds, promoting the completion of missing parts for unseen regions while preserving the local details on the input. For incorporating multi-scale pixel-aligned and voxel-aligned features, the authors use the Squeeze-and-Excitation attention to capture and fully use channel interdependencies. For the multi-view reconstruction, the proposed depth-enhanced attention explicitly excites the network to “sense” the geometric structure for a more reasonable feature aggregation. Experiments and results show that our method outperforms current RGB and depth-based SOTA methods on the challenging data from Twindom and Thuman3.0, and achieves a detailed and completed human reconstruction, balancing performance and efficiency well.

Neural machine translation (NMT) has advanced with deep learning and large-scale multilingual models, yet translating low-resource languages often lacks sufficient training data and leads to hallucinations. This often results in translated content that diverges significantly from the source text. This research proposes a refined Contrastive Decoding (CD) algorithm that dynamically adjusts weights of log probabilities from strong expert and weak amateur models to mitigate hallucinations in low-resource NMT and improve translation quality. Advanced large language NMT models, including ChatGLM and LLaMA, are fine-tuned and implemented for their superior contextual understanding and cross-lingual capabilities. The refined CD algorithm evaluates multiple candidate translations using BLEU score, semantic similarity, and Named Entity Recognition accuracy. Extensive experimental results show substantial improvements in translation quality and a significant reduction in hallucination rates. Fine-tuned models achieve higher evaluation metrics compared to baseline models and state-of-the-art models. An ablation study confirms the contributions of each methodological component and highlights the effectiveness of the refined CD algorithm and advanced models in mitigating hallucinations. Notably, the refined methodology increased the BLEU score by approximately 30% compared to baseline models.

Deep learning’s widespread dependence on large datasets raises privacy concerns due to the potential presence of sensitive information. Differential privacy stands out as a crucial method for preserving privacy, garnering significant interest for its ability to offer robust and verifiable privacy safeguards during data training. However, classic differentially private learning introduces the same level of noise into the gradients across training iterations, which affects the trade-off between model utility and privacy guarantees. To address this issue, an adaptive differential privacy mechanism was proposed in this paper, which dynamically adjusts the privacy budget at the layer-level as training progresses to resist member inference attacks. Specifically, an equal privacy budget is initially allocated to each layer. Subsequently, as training advances, the privacy budget for layers closer to the output is reduced (adding more noise), while the budget for layers closer to the input is increased. The adjustment magnitude depends on the training iterations and is automatically determined based on the iteration count. This dynamic allocation provides a simple process for adjusting privacy budgets, alleviating the burden on users to tweak parameters and ensuring that privacy preservation strategies align with training progress. Extensive experiments on five well-known datasets indicate that the proposed method outperforms competing methods in terms of accuracy and resilience against membership inference attacks.

Referring expression comprehension (REC) aims to locate a specific region in an image described by a natural language. Existing two-stage methods generate multiple candidate proposals in the first stage, followed by selecting one of these proposals as the grounding result in the second stage. Nevertheless, the number of candidate proposals generated in the first stage significantly exceeds ground truth and the recall of critical objects is inadequate, thereby enormously limiting the overall network performance. To address the above issues, the authors propose an innovative method termed Separate Non-Maximum Suppression (Sep-NMS) for two-stage REC. Particularly, Sep-NMS models information from the two stages independently and collaboratively, ultimately achieving an overall improvement in comprehension and identification of the target objects. Specifically, the authors propose a Ref-Relatedness module for filtering referent proposals rigorously, decreasing the redundancy of referent proposals. A $��{\text{CLIP}}^{†}�$ Relatedness module based on robust multimodal pre-trained encoders is built to precisely assess the relevance between language and proposals to improve the recall of critical objects. It is worth mentioning that the authors are the pioneers in utilising a multimodal pre-training model for proposal filtering in the first stage. Moreover, an Information Fusion module is designed to effectively amalgamate the multimodal information across two stages, ensuring maximum utilisation of the available information. Extensive experiments demonstrate that the approach achieves competitive performance with previous state-of-the-art methods. The datasets used are publicly available: RefCOCO, RefCOCO+: https://doi.org/10.1007/978-3-319-46475-6_5 and RefCOCOg: https://doi.org/10.1109/CVPR.2016.9.

In the field of organic synthesis, the core objective of retrosynthetic methods is to deduce possible synthetic routes and precursor molecules for complex target molecules. Traditional retrosynthetic methods, such as template-based retrosynthesis, have high accuracy and interpretability in specific types of reactions but are limited by the scope of the template library, making it difficult to adapt to new or uncommon reaction types. Moreover, sequence-to-sequence retrosynthetic prediction methods, although they enhance the flexibility of prediction, often overlook the complexity of molecular graph structures and the actual interactions between atoms, which limits the accuracy and reliability of the predictions. To address these limitations, this paper proposes a Molecular Retrosynthesis Top-k Prediction based on the Latent Generation Process (MRLGP) that uses latent variables from graph neural networks to model the generation process and produce diverse set of reactants. Utilising an encoding method based on Graphormer, the authors have also introduced topology-aware positional encoding to better capture the interactions between atomic nodes in the molecular graph structure, thereby more accurately simulating the retrosynthetic process. The MRLGP model significantly enhances the accuracy and diversity of predictions by correlating discrete latent variables with the reactant generation process and progressively constructing molecular graphs using a variational autoregressive decoder. Experimental results on benchmark datasets such as USPTO-50k, USPTO-Full, and USPTO-DIVERSE demonstrate that MRLGP outperforms baseline models on multiple Top-k evaluation metrics. Additionally, ablation experiments conducted on the USPTO-50K dataset further validate the effectiveness of the methods used in the encoder and decoder parts of the model.

To overcome the deficiencies of single-modal emotion recognition based on facial expression or bodily posture in natural scenes, a spatial guidance and temporal enhancement (SG-TE) network is proposed for facial-bodily emotion recognition. First, ResNet50, DNN and spatial ransformer models are used to capture facial texture vectors, bodily skeleton vectors and whole-body geometric vectors, and an intraframe correlation attention guidance (S-CAG) mechanism, which guides the facial texture vector and the bodily skeleton vector by the whole-body geometric vector, is designed to exploit the spatial potential emotional correlation between face and posture. Second, an interframe significant segment enhancement (T-SSE) structure is embedded into a temporal transformer to enhance high emotional intensity frame information and avoid emotional asynchrony. Finally, an adaptive weight assignment (M-AWA) strategy is constructed to realise facial-bodily fusion. The experimental results on the BabyRobot Emotion Dataset (BRED) and Context-Aware Emotion Recognition (CAER) dataset indicate that the proposed network reaches accuracies of 81.61% and 89.39%, which are 9.61% and 9.46% higher than those of the baseline network, respectively. Compared with the state-of-the-art methods, the proposed method achieves 7.73% and 20.57% higher accuracy than single-modal methods based on facial expression or bodily posture, respectively, and 2.16% higher accuracy than the dual-modal methods based on facial-bodily fusion. Therefore, the proposed method, which adaptively fuses the complementary information of face and posture, improves the quality of emotion recognition in real-world scenarios.

The success of robot-assisted pelvic fracture reduction surgery heavily relies on the accuracy of 3D/3D feature-based registration. This process involves extracting anatomical feature points from pre-operative 3D images which can be challenging because of the complex and variable structure of the pelvis. PointMLP_RegNet, a modified PointMLP, was introduced to address this issue. It retains the feature extraction module of PointMLP but replaces the classification layer with a regression layer to predict the coordinates of feature points instead of conducting regular classification. A flowchart for an automatic feature points extraction method was presented, and a series of experiments was conducted on a clinical pelvic dataset to confirm the accuracy and effectiveness of the method. PointMLP_RegNet extracted feature points more accurately, with 8 out of 10 points showing less than 4 mm errors and the remaining two less than 5 mm. Compared to PointNet++ and PointNet, it exhibited higher accuracy, robustness and space efficiency. The proposed method will improve the accuracy of anatomical feature points extraction, enhance intra-operative registration precision and facilitate the widespread clinical application of robot-assisted pelvic fracture reduction.

In recent years, there has been a concerted effort to improve anomaly detection techniques, particularly in the context of high-dimensional, distributed clinical data. Analysing patient data within clinical settings reveals a pronounced focus on refining diagnostic accuracy, personalising treatment plans, and optimising resource allocation to enhance clinical outcomes. Nonetheless, this domain faces unique challenges, such as irregular data collection, inconsistent data quality, and patient-specific structural variations. This paper proposed a novel hybrid approach that integrates heuristic and stochastic methods for anomaly detection in patient clinical data to address these challenges. The strategy combines HPO-based optimal Density-Based Spatial Clustering of Applications with Noise for clustering patient exercise data, facilitating efficient anomaly identification. Subsequently, a stochastic method based on the Interquartile Range filters unreliable data points, ensuring that medical tools and professionals receive only the most pertinent and accurate information. The primary objective of this study is to equip healthcare professionals and researchers with a robust tool for managing extensive, high-dimensional clinical datasets, enabling effective isolation and removal of aberrant data points. Furthermore, a sophisticated regression model has been developed using Automated Machine Learning (AutoML) to assess the impact of the ensemble abnormal pattern detection approach. Various statistical error estimation techniques validate the efficacy of the hybrid approach alongside AutoML. Experimental results show that implementing this innovative hybrid model on patient rehabilitation data leads to a notable enhancement in AutoML performance, with an average improvement of 0.041 in the $��R^2�$ score, surpassing the effectiveness of traditional regression models.

Existing Chinese named entity recognition (NER) research utilises 1D lexicon-based sequence labelling frameworks, which can only recognise flat entities. While lexicons serve as prior knowledge and enhance semantic information, they also pose completeness and resource requirements limitations. This paper proposes a template-based classification (TC) model to avoid lexicon issues and to identify nested entities. Template-based classification provides a template word for each entity type, which utilises contrastive learning to integrate the common characteristics among entities with the same category. Contrastive learning makes template words the centre points of their category in the vector space, thus improving generalisation ability. Additionally, TC presents a 2D table-filling label scheme that classifies entities based on the attention distribution of template words. The proposed novel decoder algorithm enables TC recognition of both flat and nested entities simultaneously. Experimental results show that TC achieves the state-of-the-art performance on five Chinese datasets.