International Journal of Pattern Recognition and Artificial Intelligence

Multi-label classification problem, where each instance can be associated with multiple labels, has received considerable attention from machine learning community. To address the inherent challenges of multi-label classification including data imbalance, label dependence, and high dimensionality, ensemble approaches have been developed, gaining popularity across various real-world applications. This paper proposes a novel ensemble method called ConfBoost that addresses these challenges and enhances the generalization ability of learning systems. ConfBoost which is a meta-model based on a weighted stacking paradigm using local confidence, combines heterogeneous and complementary ensembles of multi-label classifiers. The proposed approach achieves two main objectives: Firstly, by focusing on label weights based on their confidence scores, the model can generate more relevant predictions and enhance the accuracy at the base-level by mitigating the impact of irrelevant labels during the stacking process. Moreover, assigning higher weights to certain labels exhibits better discrimination and adaptability to capture complex label relationships. Second, applying adjusted thresholds enables the model to generate predictions adapted to the specific characteristics of each label, effectively addressing imbalanced label distributions. Extensive experiments on publicly available datasets demonstrate that ConfBoost outperforms conventional combination methods and consistently surpasses related state-of-the-art methods. These findings highlight the effectiveness and potential of ConfBoost as an advanced ensemble method for multi-label classification tasks.

The neglect of perception ability to the surrounding traffic conditions has always been the major cause of traffic accidents and the inattention to blind spots is the most important factor during driving. Existing solutions are facing the problems of using expensive equipment, wrong classification of the target object type, not suitable for nighttime, and incorrectly determining if the target object is in the blind zones. This paper aims to improve driving perception ability by developing an all-day object detection and recognition system with more accurate performance for blind zones. The proposed method uses a general-purpose camera as a single input and a two-stage deep network architecture for object detection and recognition. The proposed system is based on a two-stage cascaded network structure. At first, the style conversion process is performed to convert the daytime and nighttime images with different brightness into consistent brightness. Then the objects in the visual blind zones are detected and identified. Therefore, the accuracy of object detection can be significantly improved. Due to the diversity and complexity of Taiwan’s road conditions, the public databases cannot effectively fulfill local application model training needs. Therefore, we have built training data set from available night images. Experimental results show that the proposed method has demonstrated promising performance in all-day object detection and recognition for blind zones.

With the improvement of digital image display technology, the “secondary imaging” caused by digital cameras is also gradually popularized, and the quality of the recaptured image formed by this imaging is also getting higher and higher, and this kind of high-quality fake image has caused great threat to digital images security. We propose a neural network-based recaptured image identification algorithm and use the difference between two types of images to build the identification algorithm in the frequency domain. The algorithm uses filtering to obtain the feature images which are the high-frequency and low-frequency filtering images, in order to further distinguish the image differences, the direction of the filtered image obtained from high-frequency images, each direction of the filtered image contains high-frequency information at different angles, and the low-frequency image is downsampled. At the same time, the low-frequency image is downsampled to obtain a multi-scale filtered image. The algorithm extracts the features from previous images as the feature values for classification, and finally uses neural networks for classification to obtain the classification results, and these prove that the algorithm presented is able to differentiate the recaptured images effectively in this paper.

Classification and identification of objects are complex and challenging in pattern recognition and artificial intelligence if a shaky and nonshaky camera captures the videos at different distances during the day and nighttime. This work presents a model for classifying a given video as a static, uniform, or arbitrarily moving videos so that the complexity of the problem can be reduced. To avoid the threat of different distances between the objects and the camera, the proposed work introduces new steps for estimating the depth of the objects in the video frames. We explore locally weighted linear regression for feature extraction from depth information based on the notion that the regression line fits almost all the points for uniformity and does not fit for arbitrary moving. The extracted features are fed to a random forest classifier to classify static, uniform, or arbitrary moving video. The results on a large dataset, which includes videos captured day and night, show that the proposed method successfully classifies static, uniform and arbitrary videos with 0.86, 1.00 and 0.67 F-measures, respectively. Overall, our method obtains 87% accuracy for classification of static, uniform and arbitrary video, which is superior to the state-of-the-art methods.

With the widespread adoption of virtual reality and 360-degree video, there is a pressing need for objective metrics to assess quality in this immersive panoramic format reliably. However, existing image quality assessment models developed for traditional fixed-viewpoint content do not fully consider the specific perceptual issues involved in 360-degree viewing. This paper proposes a 360-degree image full-reference quality assessment (FR-IQA) methodology based on a multi-channel architecture. The proposed 360-degree FR-IQA method further optimizes and identifies the distorted image quality using two easily obtained useful saliency and depth-aware image features. The convolutional neural network (CNN) is designed for training. Furthermore, the proposed method accounts for predicting user viewing behaviors within 360-degree images, which will further benefit the multi-channel CNN architecture and enable the weighted average pooling of the predicted FR-IQA scores. The performance is evaluated on publicly available databases to demonstrate the advantages brought by the proposed multi-channel model in performance evaluation and cross-database evaluation experiments, where it outperforms other state-of-the-art ones. Moreover, an ablation study exhibits good generalization ability and robustness.

Channel attention has been shown to improve the performance of deep convolutional neural networks efficiently. Channel attention adaptively recalibrates the importance of each channel, determining what to attend to. However, channel attention only encodes inter-channel information but neglects the importance of positional information. Positional information is crucial in determining where to attend to. To address this issue, we propose a novel channel-spatial attention method named Spatial-Decomposition-Aggregation Attention (SDAA) method. First, a high-axis spatial direction is decomposed into multiple low-axis spatial directions. Then, a shared transformation sub-unit establishes attention in each low-axis space direction. Next, all the low-axis attention masks are aggregated into a high-axis attention mask. Finally, the generated high-axis attention mask is fused into the input features, thus enhancing the input features. Essentially, our method is a divide-and-conquer process. Experimental results demonstrate that our SDAA method outperforms the existing channel-spatial attention methods.

The task of scene text recognition involves processing information from two modalities: images and text, thereby requiring models to have the ability to extract features from images and model sequences simultaneously. Although linguistic knowledge greatly aids scene text recognition tasks, the extensive use of language models in sequence modeling and model prediction stages in recent years has made model architectures increasingly complex and inefficient. In this paper, we propose LCSTR, a pure convolutional visual model that can complete text recognition without the need for attention mechanisms or language models. This approach applies large kernels to text recognition tasks for the first time, extracting word-level text information through large text-aware blocks, capturing long-range dependencies between characters, and using small text-aware blocks to obtain local features within characters. Experiments show that this model strikes a good trade-off between accuracy and speed, achieving notable results on seven public benchmarks, validating the generalizability and effectiveness of this method. Furthermore, owing to the absence of a language module, this model demonstrates remarkable accuracy even in limited sample scenarios, and the lightweight and low computational overhead features make it suitable for engineering applications.

Deep learning-based video coding methods have demonstrated superior performance compared to classical video coding standards in recent years. The vast majority of the existing deep video coding (DVC) networks are based on convolutional neural networks (CNNs), and their main drawback is that since CNNs are affected by the size of the receptive field, they cannot effectively handle long-range dependencies and local detail recovery. Therefore, how to better capture and process the overall structure as well as local texture information in the video coding task is the core issue. Notably, the transformer employs a self-attention mechanism that captures dependencies between any two positions in the input sequence without being constrained by distance limitations. This is an effective solution to the problem described above. In this paper, we propose end-to-end transformer-based adaptive video coding (TAVC). First, we compress the motion vector and residuals through a compression network built on the vision transformer (ViT) and design the motion compensation network based on ViT. Second, based on the requirement of video coding to adapt to different resolution inputs, we introduce a position encoding generator (PEG) as adaptive position encoding (APE) to maintain its translation invariance across different resolution video coding tasks. The experiment shows that for multiscale structural similarity index measurement (MS-SSIM) metrics, this method exhibits significant performance gaps compared to conventional engineering codecs, such as $\times 264$, $\times 265$, and VTM-15.2. We also achieved a good performance improvement compared to the CNN-based DVC methods. In the case of peak signal-to-noise ratio (PSNR) evaluation metrics, TAVC also achieves good performance.

In this paper, a parallel decoder and a word group prediction module are proposed to speed up decoding and improve the effect of captions. The features of the image extracted by the encoder are linearly projected to different word groups, and then a unique relaxed mask matrix is designed to improve the decoding speed and the caption effect. First, since image captioning is composed of many words, sentences can also be broken down into word groups or words according to their syntactic structure, and we achieve this function through constituency parsing. Second, we make full use of the extracted features to predict the size of word groups. Then, a new embedding representing the information of the word is proposed based on word embedding. Finally, with the help of word groups, we design a mask matrix to modify the decoding process so that each step of the model can produce one or more words in parallel. Experiments on public datasets demonstrate that our method can reduce the time complexity while maintaining competitive performance.

In recent years, multi-stream gaze estimation methods have become mainstream, which estimate gaze point by eye picture or combine with facial appearance, have achieved considerable accuracy. However, these methods based on a single camera fail to obtain accurate eye spatial position information. To address this issue, we propose a multi-stream gaze estimation model that incorporates spatial position information. We acquire eye spatial position information using a stereo camera and fuse eye image features with eye spatial position information using a ResNet network with a fused attention mechanism. Additionally, we perform calibration of eye image features using the computed eye spatial position information. Our model demonstrates superior performance on our experimental dataset.

The timely and precise discovery of traffic signs is considered an effective part of modeling automated vehicle driving. However, the dimension of traffic signs accounted for a lower ratio of input pictures which elevated the complexity of discovery. Hence, a new model is devised using faster region-based convolution neural network (faster R-CNN) traffic for detecting traffic signs. The Region of Interest (RoI) extraction and Median filter are executed for discarding the superfluous data from the dataset. The method extracted a Pyramid Histogram of Oriented Gradient (PHoG), local vector pattern (LVP), CNN and ResNet features for generating beneficial information. It is used to lessen the loss of contextual data and the data augmentation is further applied for making the training of the model more stable and time-saving. The traffic sign recognition is executed with faster R-CNN wherein the tuning of hyperparameters such as batch normalization rate, epoch and learning rate is determined by the proposed pelican cuckoo search (PCS). The method revealed improved efficacy without presenting additional computational costs in the model. Moreover, the faster R-CNN is termed the finest technique to enhance the discovery of traffic signs. The proposed PCS-based faster R-CNN outperformed with the highest precision 92.7%, specificity of 93.7% and $F$-measure of 93.2%.

Occlusive keypoints has been a challenge for human pose estimation, especially the mutual occlusion of human bodies. One possible solution to this problem is to utilize multi-scale features, where small scale features are capable of identifying keypoints, while large-scale features can capture the relationship between keypoints. Feature fusion among multi-scale features allows for the exchange of information between keypoints, facilitating the inference of occluded keypoints based on the identified keypoints. However, it’s found that there are invalid features in feature fusion which will interfere valid feature. In this paper, we propose multi-scale feature refined network (MSFRNet) based on HRNet and a new attention module namely multi-resolution attention module (MRAM). The proposed MRAM is designed to strengthen the effective information while suppressing redundant information. It has multiple inputs and outputs and can learn the relationships between keypoints while retaining detailed information. The proposed MSFRNet outperforms HRNet, achieving a 1.4AP improvement on the COCO dataset with only a marginal computational increase of 0.35 GFLOPs. Additionally, it demonstrates superior performance with a 0.9AP, 0.7AP, and 1.8AP improvement on the MPII, CrowdPose and OCHuman datasets, respectively. Furthermore, compared with the latest attention mechanism PSA, the MSFRNet exhibits lower computational cost while maintaining the same pose-estimation accuracy.

Presently, one of the foremost health issues and an extremely transferrable disease is Tuberculosis which is spreading worldwide. Tuberculosis is generally produced by mycobacterium tuberculosis and can cause death if it is not detected at premature stages. Therefore, a precise and efficient approach is essential for the identification of tuberculosis. The physical analysis of sputum smears through fluorescence microscopy for the detection process and also the treatment process is a subjective and difficult task. Thus, the hybrid optimization-driven deep learning approach is designed in this paper for infection-level identification of severity prediction of tuberculosis. Here, a sputum image is considered for performing tuberculosis detection and it is acquired from a database. For executing bacilli segmentation, the Deep Contour Aware Network (DCAN) model is used. In addition, DCAN is trained by the introduced Fractional Artificial Hummingbird Algorithm (FAHA). Additionally, the LeNet scheme is applied to detect the infection level and severity detection. The LeNet is also trained using Coot-FAHA thereby; the system performance is highly increased. The devised optimization-enabled deep learning method achieved improved performance with precision, F1-score, and recall of 0.9326, 0.9386, and 0.9163, respectively.

Deep learning-based algorithms for detecting objects in remote sensing images have produced excellent results recently. However, the target recognition and classification process of remote sensing images has problems such as dense targets, uneven distribution, large-scale changes and complex backgrounds. In order to improve the effectiveness of existing detection methods, based on the YOLOX algorithm, a remote sensing image object detection algorithm introducing Global Attention Mechanism (GAM) and Feature Enhancement Module (FEM) proposed, named the M²-YOLOX(GAM+FEM+YOLOX) algorithm. First, a novel GAM module is developed that employs a sequential channel-space attention mechanism and redesigns the Convolutional Block Attention Module (CBAM), to address the issues of low effective information extraction and weak information representation of the feature map in the backbone network. CBAM is capable of amplifying global dimensional interaction features while reducing information dispersion as well. Second, the goal of the FEM is to improve the target feature extraction capabilities of the backbone feature extraction network by fusing numerous perceptual field features in lower-level feature maps. Then, the Flexible Rectified Linear Unit (FReLU) activation function is introduced under the action of feature fusion and global attention mechanism. Four-way feature map output in Feature Pyramid Networks (FPN) with Non-Maximum Suppression (NMS) and score filtering for object detection and output results. In comparison to the YOLOX algorithm, the experimental results show that the mAP value of the M²-YOLOX algorithm is improved by 0.0123, the LAMR value is decreased by 0.0150, the Precision rate value is increased by 0.105450, the Recall value is increased by 0.053250, and the mF1 value is increased by 0.0425.