CAAI Transactions on Intelligence Technology最新文献

In recent years, the transformer model has demonstrated excellent performance in computer vision (CV) applications. The key lies in its guided representation attention mechanism, which uses dot-product to depict complex feature relationships, and comprehensively understands the context semantics to obtain feature weights. Then feature enhancement is implemented by guiding the target matrix through feature weights. However, the uncertainty and inconsistency of features are widespread that prone to confusion in the description of relationships within dot-product attention mechanisms. To solve this problem, this paper proposed a novel approximate-guided representation learning methodology for vision transformer. The kernelised matroids fuzzy rough set is defined, wherein the closed sets inside kernelised fuzzy information granules of matroids structures can constitute the subspace of lower approximation in rough sets. Thus, the kernel relation is employed to characterise image feature granules that will be reconstructed according to the independent set in matroids theory. Then, according to the characteristics of the closed set within matroids, the feature attention weight is formed by using the lower approximation to realise the approximate guidance of features. The approximate-guided representation mechanism can be flexibly deployed as a plug-and-play component in a wide range of CV tasks. Extensive empirical results demonstrate that the proposed method outperforms the majority of advanced prevalent models, especially in terms of robustness.

Medical images play a crucial role in diagnosis, treatment procedures and overall healthcare. Nevertheless, they also pose substantial risks to patient confidentiality and safety. Safeguarding the confidentiality of patients' data has become an urgent and practical concern. We present a novel approach for reversible data hiding for colour medical images. In a hybrid domain, we employ AlexNet, tuned with watershed transform (WST) and L-shaped fractal Tromino encryption. Our approach commences by constructing the host image's feature vector using a pre-trained AlexNet model. Next, we use the watershed transform to convert the extracted feature vector into a vector for a topographic map, which we then encrypt using an L-shaped fractal Tromino cryptosystem. We embed the secret image in the transformed image vector using a histogram-based embedding strategy to enhance payload and visual fidelity. When there are no attacks, the RDHNet exhibits robust performance, can be reversed to the original image and maintains a visually appealing stego image, with an average PSNR of 73.14 dB, an SSIM of 0.9999 and perfect values of NC = 1 and BER = 0 under normal conditions. The proposed RDHNet demonstrates a robust ability to withstand detrimental geometric and noise-adding attacks as well as various steganalysis methods. Furthermore, our RDHNet method initiative demonstrates efficacy in tackling contemporary confidentiality issues.

In recent years, 3D object detection using neural radiance fields (NeRF) has advanced significantly, yet challenges remain in effectively utilising the density field. Current methods often treat NeRF as a geometry learning tool or rely on volume rendering, neglecting the density field's potential and feature dependencies. To address this, we propose NeRF-C3D, a novel framework incorporating a multi-scale feature fusion module with channel attention (MFCA). MFCA leverages channel attention to model feature dependencies, dynamically adjusting channel weights during fusion to enhance important features and suppress redundancy. This optimises density field representation and improves feature discriminability. Experiments on 3D-FRONT, Hypersim, and ScanNet demonstrate NeRF-C3D's superior performance validating MFCA's effectiveness in capturing feature relationships and showcasing its innovation in NeRF-based 3D detection.

Lumbar degenerative disc diseases constitute a major contributor to lower back pain. In pursuit of an enhanced understanding of lumbar degenerative pathology and the development of more effective treatment modalities, the application of precise measurement techniques for lumbar segment kinematics is imperative. This study aims to pioneer a novel automated lumbar spine orientation estimation method using deep learning techniques, to facilitate the automatic 2D–3D pre-registration of the lumbar spine during physiological movements, to enhance the efficiency of image registration and the accuracy of spinal segment kinematic measurements. A total of 12 asymptomatic volunteers were enrolled and captured in 2 oblique views with 7 different postures. Images were used for deep learning model development training and evaluation. The model was composed of a segmentation module using Mask R-CNN and an estimation module using ResNet50 architecture with a Squeeze-and-Excitation module. The cosine value of the angle between the prediction vector and the vector of ground truth was used to quantify the model performance. Data from another two prospective recruited asymptomatic volunteers were used to compare the time cost between model-assisted registration and manual registration without a model. The cosine values of vector deviation angles at three axes in the cartesian coordinate system were 0.9667 ± 0.004, 0.9593 ± 0.0047 and 0.9828 ± 0.0025, respectively. The value of the angular deviation between the intermediate vector obtained by utilising the three direction vectors and ground truth was 10.7103 ± 0.7466. Results show the consistency and reliability of the model's predictions across different experiments and axes and demonstrate that our approach significantly reduces the registration time (3.47 ± 0.90 min vs. 8.10 ± 1.60 min, p < 0.001), enhances the efficiency, and expands its broader utilisation of clinical research about kinematic measurements.

Lung cancer (LC) is a major cancer which accounts for higher mortality rates worldwide. Doctors utilise many imaging modalities for identifying lung tumours and their severity in earlier stages. Nowadays, machine learning (ML) and deep learning (DL) methodologies are utilised for the robust detection and prediction of lung tumours. Recently, multi modal imaging emerged as a robust technique for lung tumour detection by combining various imaging features. To cope with that, we propose a novel multi modal imaging technique named versatile scale malleable image integration and patch wise attention network ($��{\text{VSMI}}^2�-�\text{PANet}�$) which adopts three imaging modalities named computed tomography (CT), magnetic resonance imaging (MRI) and single photon emission computed tomography (SPECT). The designed model accepts input from CT and MRI images and passes it to the $��{\text{VSMI}}^2�$ module that is composed of three sub-modules named image cropping module, scale malleable convolution layer (SMCL) and PANet module. CT and MRI images are subjected to image cropping module in a parallel manner to crop the meaningful image patches and provide them to the SMCL module. The SMCL module is composed of adaptive convolutional layers that investigate those patches in a parallel manner by preserving the spatial information. The output from the SMCL is then fused and provided to the PANet module. The PANet module examines the fused patches by analysing its height, width and channels of the image patch. As a result, it provides an output as high-resolution spatial attention maps indicating the location of suspicious tumours. The high-resolution spatial attention maps are then provided as an input to the backbone module which uses light wave transformer (LWT) for segmenting the lung tumours into three classes, such as normal, benign and malignant. In addition, the LWT also accepts SPECT image as input for capturing the variations precisely to segment the lung tumours. The performance of the proposed model is validated using several performance metrics, such as accuracy, precision, recall, F1-score and AUC curve, and the results show that the proposed work outperforms the existing approaches.

In recent years, camouflage technology has evolved from single-spectral-band applications to multifunctional and multispectral implementations. Hyperspectral imaging has emerged as a powerful technique for target identification due to its capacity to capture both spectral and spatial information. The advancement of imaging spectroscopy technology has significantly enhanced reconnaissance capabilities, offering substantial advantages in camouflaged target classification and detection. However, the increasing spectral similarity between camouflaged targets and their backgrounds has significantly compromised detection performance in specific scenarios. Conventional feature extraction methods are often limited to single, shallow spectral or spatial features, failing to extract deep features and consequently yielding suboptimal classification accuracy. To address these limitations, this study proposes an innovative 3D-2D convolutional neural networks architecture incorporating depthwise separable convolution (DSC) and attention mechanisms (AM). The framework first applies dimensionality reduction to hyperspectral images and extracts preliminary spectral-spatial features. It then employs an alternating combination of 3D and 2D convolutions for deep feature extraction. For target classification, the LogSoftmax function is implemented. The integration of depthwise separable convolution not only enhances classification accuracy but also substantially reduces model parameters. Furthermore, the attention mechanisms significantly improve the network's ability to represent multidimensional features. Extensive experiments were conducted on a custom land-based hyperspectral image dataset. The results demonstrate remarkable classification accuracy: 98.74% for grassland camouflage, 99.13% for dead leaf camouflage and 98.94% for wild grass camouflage. Comparative analysis shows that the proposed framework is outstanding in terms of classification accuracy and robustness for camouflage target classification.

In clinical diagnosis, magnetic resonance imaging (MRI) allows different contrast images to be obtained. High-resolution (HR) MRI presents fine anatomical structures, which is important for improving the efficiency of expert diagnosis and realising smart healthcare. However, due to the cost of scanning equipment and the time required for scanning, obtaining an HR brain MRI is quite challenging. Therefore, to improve the quality of images, reference-based super-resolution technology has come into existence. Nevertheless, the existing methods still have some drawbacks: (1) The advantages of different contrast images are not fully utilised. (2) The slice-by-slice scanning nature of magnetic resonance imaging is not considered. (3) The ability to capture contextual information and to match and fuse multi-scale, multi-contrast features is lacking. In this paper, we propose the multi-slice aware matching and fusion (MSAMF) network, which makes full use of multi-slice reference images information by introducing a multi-slice aware module and multi-scale matching strategy to capture corresponding contextual information in reference features at other scales. To further integrate matching features, a multi-scale fusion mechanism is also designed to progressively fuse multi-scale matching features, thereby generating more detailed super-resolution images. The experimental results support the benefits of our network in enhancing the quality of brain MRI reconstruction.

With the increasing constraints of hardware devices, there is a growing demand for compact models to be deployed on device endpoints. Knowledge distillation, a widely used technique for model compression and knowledge transfer, has gained significant attention in recent years. However, traditional distillation approaches compare the knowledge of individual samples indirectly through class prototypes overlooking the structural relationships between samples. Although recent distillation methods based on contrastive learning can capture relational knowledge, their relational constraints often distort the positional information of the samples leading to compromised performance in the distilled model. To address these challenges and further enhance the performance of compact models, we propose a novel approach, termed contrastive learning-based multi-level knowledge distillation (CLMKD). The CLMKD framework introduces three key modules: class-guided contrastive distillation, gradient relation contrastive distillation, and semantic similarity distillation. These modules are effectively integrated into a unified framework to extract feature knowledge from multiple levels, capturing not only the representational consistency of individual samples but also their higher-order structure and semantic similarity. We evaluate the proposed CLMKD method on multiple image classification datasets and the results demonstrate its superior performance compared to state-of-the-art knowledge distillation methods.

To address the challenges of jerky movements and poor tracking performance in outdoor environments for a following-type mobile robot, a novel marker-based human–machine following-motion control strategy is explored. This strategy decouples the control of linear velocity and angular velocity, handling them separately. First, in the design of linear-velocity control, using the identification of markers to determine the distance between the human and the robot, an enhanced virtual spring model is developed. This involves designing a weighted dynamic damping coefficient to address the rationality issues of the range and trend of the robot's following speed, thereby improving its smoothness and reducing the risk of target loss. Second, in the design of angular velocity control, a new concept of an ‘insensitive zone’ based on the offset of the marker's centre point is proposed, combined with a fuzzy controller to address the issue of robot jitter and enhance its resistance to interference. The experimental results indicate that the average variance in the human–robot distance is 1.037 m, whereas the average variance in the robot's linear velocity is 0.345 m/s. Due to the design of an insensitive region in parameter-adaptive fuzzy control, the average variance of angular velocity is only 0.031 rad/s. When the human–robot distance exhibits significant fluctuations, the fluctuations in both linear and angular velocities are comparatively small, allowing for stable and smooth following movements. This demonstrates the effectiveness of the motion control strategy designed in this study.

The challenge of optimising multimodal functions within high-dimensional domains constitutes a notable difficulty in evolutionary computation research. Addressing this issue, this study introduces the Deep Backtracking Bare-Bones Particle Swarm Optimisation (DBPSO) algorithm, an innovative approach built upon the integration of the Deep Memory Storage Mechanism (DMSM) and the Dynamic Memory Activation Strategy (DMAS). The DMSM enhances the memory retention for the globally optimal particle, promoting interaction between standard particles and their historically optimal counterparts. In parallel, DMAS assures the updated position of the globally optimal particle is appropriately aligned with the deep memory repository. The efficacy of DBPSO was rigorously assessed through a series of simulations employing the CEC2017 benchmark suite. A comparative analysis juxtaposed DBPSO's performance against five contemporary evolutionary algorithms across two experimental conditions: Dimension-50 and Dimension-100. In the 50D trials, DBPSO attained an average ranking of 2.03, whereas in the 100D scenarios, it improved to an average ranking of 1.9. Further examination utilising the CEC2019 benchmark functions revealed DBPSO's robustness, securing four first-place finishes, three second-place standings, and three third-place positions, culminating in an unmatched average ranking of 1.9 across all algorithms. These empirical results corroborate DBPSO's proficiency in delivering precise solutions for complex, high-dimensional optimisation challenges.