CAAI Transactions on Intelligence Technology最新文献

Neural image compression (NIC) has shown remarkable rate-distortion (R-D) efficiency. However, the considerable computational and spatial complexity of most NIC methods presents deployment challenges on resource-constrained devices. We introduce a lightweight neural image compression framework designed to efficiently process both local and global information. In this framework, the convolutional branch extracts local information, whereas the frequency domain branch extracts global information. To capture global information without the high computational costs of dense pixel operations, such as attention mechanisms, Fourier transform is employed. This approach allows for the manipulation of global information in the frequency domain. Additionally, we employ feature shift operations as a strategy to acquire large receptive fields without any computational cost, thus circumventing the need for large kernel convolution. Our framework achieves a superior balance between rate-distortion performance and complexity. On varying resolution sets, our method not only achieves rate-distortion (R-D) performance on par with versatile video coding (VVC) intra and other state-of-the-art (SOTA) NIC methods but also exhibits the lowest computational requirements, with approximately 200 KMACs/pixel. The code will be available at https://github.com/baoyu2020/SFNIC.

The 100-kernel weight of corn seed is a crucial metric for assessing corn quality, and the current measurement means mostly involve manual counting of kernels followed by weighing on a balance, which is labour-intensive and time-consuming. Aiming to address the problem of low efficiency in measuring the 100-kernel weight of corn seeds, this study proposes a measurement method based on deep learning and machine vision. In this study, high-contrast camera technology was utilised to capture image data of corn seeds. And improvements were made to the feature extraction network of the YOLOv5 model by incorporating the MobileNetV3 network structure. The novel model employs deep separable convolution to decrease parameters and computational load. It incorporates a linear bottleneck and inverted residual structure to enhance efficiency. It introduces an SE attention mechanism for direct learning of channel number features and updates the activation function. Algorithms and experiments were subsequently designed to calculate the 100-grain weight in conjunction with the output of the model. The outcomes revealed that the enhanced model in this study achieved an accuracy of 90.1%, a recall rate of 91.3%, and a mAP (mean average precision) value of 92.2%. While meeting production requirements, this model significantly reduces the number of parameters compared to alternative models—50% of the original model. In an applied study focused on measuring the 100-kernel weight of corn seeds, the counting accuracy yielded a remarkable 97.18%, while the accuracy for weight measurement results reached 94.2%. This study achieves both efficient and precise measurement of the 100-kernel weight of maize seeds, presenting a novel perspective in the exploration of maize seed weight.

Tibetan medical named entity recognition (Tibetan MNER) involves extracting specific types of medical entities from unstructured Tibetan medical texts. Tibetan MNER provide important data support for the work related to Tibetan medicine. However, existing Tibetan MNER methods often struggle to comprehensively capture multi-level semantic information, failing to sufficiently extract multi-granularity features and effectively filter out irrelevant information, which ultimately impacts the accuracy of entity recognition. This paper proposes an improved embedding representation method called syllable–word–sentence embedding. By leveraging features at different granularities and using un-scaled dot-product attention to focus on key features for feature fusion, the syllable–word–sentence embedding is integrated into the transformer, enhancing the specificity and diversity of feature representations. The model leverages multi-level and multi-granularity semantic information, thereby improving the performance of Tibetan MNER. We evaluate our proposed model on datasets from various domains. The results indicate that the model effectively identified three types of entities in the Tibetan news dataset we constructed, achieving an F1 score of 93.59%, which represents an improvement of 1.24% compared to the vanilla FLAT. Additionally, results from the Tibetan medical dataset we developed show that it is effective in identifying five kinds of medical entities, with an F1 score of 71.39%, which is a 1.34% improvement over the vanilla FLAT.

Visual perception is critical in robotic operations, particularly in collaborative and autonomous robot systems. Through efficient visual systems, robots can acquire and process environmental information in real-time, recognise objects, assess spatial relationships, and make adaptive decisions. This review aims to provide a comprehensive overview of the latest advancements in the field of vision as applied to robotic perception, focusing primarily on visual applications in the areas of object perception, self-perception, human–robot collaboration, and multi-robot collaboration. By summarising the current state of development and analysing the challenges and opportunities that remain in these areas, this paper offers a thorough examination of the integration of visual perception with operational robotics. It further inspires future research and drives the application and development of visual perception across various robotic domains, enabling operational robots to better adapt to complex environments and reliably accomplish tasks.

Selecting appropriate tourist attractions to visit in real time is an important problem for travellers. Since recommenders proactively suggest items based on user preference, they are a promising solution for this problem. Travellers visit tourist attractions sequentially by considering multiple attributes at the same time. Therefore, it is desirable to consider this when developing recommenders for tourist attractions. Using GRU4REC, we proposed RNN-based sequence-aware recommenders (RNN-SARs) that use multiple sequence datasets for training the recommended model, named multi-RNN-SARs. We proposed two types of multi-RNN-SARs—concatenate-RNN-SARs and parallel-RNN-SARs. In order to evaluate multi-RNN-SARs, we compared hit rate (HR) and mean reciprocal rank (MRR) of the item-based collaborative filtering recommender (item-CFR), RNN-SAR with the single-sequence dataset (basic-RNN-SAR), multi-RNN-SARs and the state-of-the-art SARs using a real-world travel dataset. Our research shows that multi-RNN-SARs have significantly higher performances compared to item-CFR. Not all multi-RNN-SARs outperform basic-RNN-SAR but the best multi-RNN-SAR achieves comparable performance to that of the state-of-the-art algorithms. These results highlight the importance of using multiple sequence datasets in RNN-SARs and the importance of choosing appropriate sequence datasets and learning methods for implementing multi-RNN-SARs in practice.

Ensemble learning, a pivotal branch of machine learning, amalgamates multiple base models to enhance the overarching performance of predictive models, capitalising on the diversity and collective wisdom of the ensemble to surpass individual models and mitigate overfitting. In this review, a four-layer research framework is established for the research of ensemble learning, which can offer a comprehensive and structured review of ensemble learning from bottom to top. Firstly, this survey commences by introducing fundamental ensemble learning techniques, including bagging, boosting, and stacking, while also exploring the ensemble's diversity. Then, deep ensemble learning and semi-supervised ensemble learning are studied in detail. Furthermore, the utilisation of ensemble learning techniques to navigate challenging datasets, such as imbalanced and high-dimensional data, is discussed. The application of ensemble learning techniques across various research domains, including healthcare, transportation, finance, manufacturing, and the Internet, is also examined. The survey concludes by discussing challenges intrinsic to ensemble learning.

Prediction of the age of each individual is possible using the changing pattern of DNA methylation with age. In this paper an age prediction approach to work out multivariate regression problems using DNA methylation data is developed. In this research study a convolutional neural network (CNN)-based model optimised by the genetic algorithm (GA) is addressed. This paper contributes to enhancing age prediction as a regression problem using a union of two CNNs and exchanging knowledge between them. This specifically re-starts the training process from a possibly higher-quality point in different iterations and, consequently, causes potentially yeilds better results at each iteration. The method proposed, which is called cooperative deep neural network (Co-DeepNet), is tested on two types of age prediction problems. Sixteen datasets containing 1899 healthy blood samples and nine datasets containing 2395 diseased blood samples are employed to examine the method's efficiency. As a result, the mean absolute deviation (MAD) is 1.49 and 3.61 years for training and testing data, respectively, when the healthy data is tested. The diseased blood data show MAD results of 3.81 and 5.43 years for training and testing data, respectively. The results of the Co-DeepNet are compared with six other methods proposed in previous studies and a single CNN using four prediction accuracy measurements (R², MAD, MSE and RMSE). The effectiveness of the Co-DeepNet and superiority of its results is proved through the statistical analysis.

Nonlinear transforms have significantly advanced learned image compression (LIC), particularly using residual blocks. This transform enhances the nonlinear expression ability and obtain compact feature representation by enlarging the receptive field, which indicates how the convolution process extracts features in a high dimensional feature space. However, its functionality is restricted to the spatial dimension and network depth, limiting further improvements in network performance due to insufficient information interaction and representation. Crucially, the potential of high dimensional feature space in the channel dimension and the exploration of network width/resolution remain largely untapped. In this paper, we consider nonlinear transforms from the perspective of feature space, defining high-dimensional feature spaces in different dimensions and investigating the specific effects. Firstly, we introduce the dimension increasing and decreasing transforms in both channel and spatial dimensions to obtain high dimensional feature space and achieve better feature extraction. Secondly, we design a channel-spatial fusion residual transform (CSR), which incorporates multi-dimensional transforms for a more effective representation. Furthermore, we simplify the proposed fusion transform to obtain a slim architecture (CSR-sm), balancing network complexity and compression performance. Finally, we build the overall network with stacked CSR transforms to achieve better compression and reconstruction. Experimental results demonstrate that the proposed method can achieve superior rate-distortion performance compared to the existing LIC methods and traditional codecs. Specifically, our proposed method achieves 9.38% BD-rate reduction over VVC on Kodak dataset.

Recently, numerous estimation issues have been solved due to the developments in data-driven artificial neural networks (ANN) and graph neural networks (GNN). The primary limitation of previous methodologies has been the dependence on data that can be structured in a grid format. However, physiological recordings often exhibit irregular and unordered patterns, posing a significant challenge in conceptualising them as matrices. As a result, GNNs which comprise interactive nodes connected by edges whose weights are defined by anatomical junctions or temporal relationships have received a lot of consideration by leveraging implicit data that exists in a biological system. Additionally, our study incorporates a structural GNN to effectively differentiate between different degrees of infection in both the left and right hemispheres of the brain. Subsequently, demographic data are included, and a multi-task learning architecture is devised, integrating classification and regression tasks. The trials used an authentic dataset, including 800 brain x-ray pictures, consisting of 560 instances classified as moderate cases and 240 instances classified as severe cases. Based on empirical evidence, our methodology demonstrates superior performance in classification, surpassing other comparison methods with a notable achievement of 92.27% in terms of area under the curve as well as a correlation coefficient of 0.62.