Information Technology and Control

Learning Sliding Policy of Flat Multi-target Objects in Clutter Scenes

Liangdong Wu — 2024-03-22

In clutter scenes, one or several targets need to be obtained, which is hard for robot manipulation task. Especially, when the targets are flat objects like book, plates, due to limitation of common robot end-effectors, it will be more challenging. By employing pre-grasp operation like sliding, it becomes feasible to rearrange objects and shift the target towards table edge, enabling the robot to grasp it from a lateral perspective. In this paper, the proposed method transfers the task into a Parameterized Action Markov Decision Process to solve the problem, which is based on deep reinforcement learning. The mask images are taken as one of observations to the network for avoiding the impact of noise of original image. In order to improve data utilization, the policy
network predicts the parameters for the sliding primitive of each object, which is weight-sharing, and then the Q-network selects the optimal execution target. Meanwhile, extra reward mechanism is adopted for improving the efficiency of task actions to cope with multiple targets. In addition, an adaptive policy scaling algorithm is proposed to improve the speed and adaptability of policy training. In both simulation and real system, our method achieves a higher task success rate and requires fewer actions to accomplish the flat multi-target sliding manipulation task within clutter scene, which verifies the effectiveness of ours.

Classification of Medicinal Plant Leaves for Types and Diseases with Hybrid Deep Learning Methods

Kiyas Kayaalp — 2024-03-22

Leaf images are often used to detect plant diseases because most disease symptoms appear on the leaves. Analyzes performed by experts in the laboratory environment are expensive and time consuming. Therefore, there is a need for automated plant disease detection systems that are both economical and can help diagnose early symptoms more accurately. In this study, a deep learning-based methodology is presented for the classification of leaf diseases of plants, which are very similar in color, texture, vein and shape and cannot be noticed by non-experts, which are important for traditional medicine and pharmaceutical industry. In the model development process, 7 pre-learning deep learning algorithms and an image data set created from plant leaves in 10 categories were preferred. The proposed model classifies the plant type and diseased condition in the dataset. In the first step of training the model, different learning rates were tested with optimum hyperparameters. In the second part, a test accuracy rate of 98.69% was achieved with the DenseNet121 model, with increased data. At the last stage, after the edge detection processes, the test accuracy value of 67.92% was reached with the DenseNet 121 model.

An Anomaly Detection Approach Based on Bidirectional Temporal Convolutional Network and Multi-Head Attention Mechanism

Rui Wang — 2024-03-22

Anomaly detection aims at detecting the data instances that deviate from the majority of data, and it is widely used in various fields for its ability to ensure the quality of the overall data. However, traditional anomaly detection methods face the problems such as low efficiency due to high data complexity and lack of data labels. At the same time, most methods only learn the forward features of time-series data, while lacking attention to the reverse features. For these disadvantages, this paper designs an anomaly detection approach called BiTCN-MHA based on the bidirectional temporal convolutional network (BiTCN) and multi-head attention (MHA) mechanism, which learns the features of anomalous data by capturing the forward and reverse temporal features in the time-series data, as well as solves the problems of feature information overload and neuron “death” by using MHA mechanism and ELU activation function, respectively, thereby quickly and accurately detecting anomalous data. Extensive experiments on six public datasets show that compared with eight state-of-the-arts, the proposed BiTCN-MHA method can improve the precision, recall, AUC and F1-Score by about 6.10%, 10.16%, 4.06% and 8.50%, respectively, especially having better detection performance on small time-series data.

Enhanced Two-Stream Bayesian Hyper Parameter Optimized 3D-CNN Inception-v3 Based Drop-ConvLSTM2D Deep Learning Model for Human Action Recognition

A. Jeyanthi — 2024-03-22

Human Action Recognition (HAR) has grown to be the toughest and most attractive concern in the domains of computer vision, communication between a person and the surroundings, and video surveillance. In variation to the conventional methods that usually make use of the Long Short Term Memory model (LSTM) for training, this work designed dropout variant Drop-ConvLSTM2D, to provide more effectiveness in regularization for deep Convolution Neural Networks (CNNs).In addition, to speed up the runtime performance of the Deep Learning model, Bayesian Hyper Parameter Optimization (BHPO) is also introduced to autonomously optimize, the hyperparameters of the trained architecture. In this study, a two-stream Bayesian Hyper Parameter optimized Drop-ConvLSTM2D model is designed for HAR to overcome the current research deficiencies. In one stream, an Inception-v3 model extracts the temporal characteristics from the optical frames which are generated through the dense flow process. In another stream, a 3D-CNN involves the mining of the spatial-temporal characteristics from the RGB frames. Finally, the features of Inception-v3 and 3D-CNN are fused using which the Drop-ConvLSTM2D model is trained to recognize human behavior. On perceptive public video datasets UCF-101, and HMDB51, the quantitative assessments are conducted on the Drop-ConvLSTM2D BHPO model. For all hyperparameters, the built model explicitly obtains optimized values in this process, which can save time and improve performance. The experimental outcome shows that with a precision of at least 3%, the designed model beats the traditional two-stream model.

An OBB Detection Algorithm of Maintenance Components Based on YOLOv5-OBB-CR

Zhexue Ge — 2024-03-22

By detecting the position of maintenance components in real-time, maintenance guidance information can be superimposed and important operational guidance can be provided for maintenance personnel. The YOLOv5-OBB-CR real-time detection algorithm is proposed for maintenance component with orientation bounding box based on improved YOLOv5-OBB. The C3 module in the original network is improved to CReToNeXt, which can more effectively enhance the network's ability to learn image features. Considering that the network learning is the labeled rotation box information, the original Loss function CIoU is improved to SIoU with angle loss information, and the improved Loss function can more effectively describe the regression of the target box. The demonstration shows that the mAP@.5 0.95 of YOLOv5-OBB-CR-s (SIoU) is 85.6%, which is 6.7% higher than the original YOLOv5 OBB algorithm.

An Estimation of Distribution Based Algorithm for Continuous Distributed Constraint Optimization Problems

Meifeng Shi — 2024-03-22

Continuous Distributed Constraint Optimization Problem(C-DCOP) is a constraint processing framework for continuous variables problems in multi-agent systems. There is a constraint cost function between two mutually restrictive agents in C-DCOP. The goal of the C-DCOP solving algorithm is to keep the sum of constraint cost functions in an extreme state. In a C-DCOP, each function is defined by a set of continuous variables. At present, some C-DCOP solving algorithms have been proposed, but there are some common problems such as the limitation of constraints cost function form, easy to fall into local optimum, and lack of anytime attribute. Aiming at these thorny problems, we propose a parallel optimization algorithm named Estimation of Distribution Based Algorithm for Continuous Distributed Constraint Optimization Problems (EDA-CD). In EDA-CD, each solution is regarded as an individual, and the distribution of agent value is jointly described by all outstanding individuals. Firstly, all agents cooperate to hold a distributed population. Secondly, each agent calculates the mean and variance of its variables to build probability models in parallel. Finally, the agent evaluates the fitness of samples and updates the probability model through cooperative communication on Breadth First Search (BFS) pseudo-tree. We theoretically prove that EDA-CD is an anytime algorithm. The extensive experimental results on four types of benchmark problems show that the proposed EDA-CD outperforms the state-of-the-art C-DCOP algorithms and has about 20% improvement in solution quality.

TSIC-CLIP: Traffic Scene Image Captioning Model Based on Clip

Hao Zhang — 2024-03-22

Image captioning in traffic scenes presents several challenges, including imprecise caption generation, lack of personalization, and an unwieldy number of model parameters. We propose a new image captioning model for traffic scenes to address these issues. The model incorporates an adapter-based fine-tuned feature extraction part to enhance personalization and a caption generation module using global weighted attention pooling to reduce model parameters and improve accuracy. The proposed model consists of four main stages. In the first stage, the Image-Encoder extracts the global features of the input image and divides it into nine sub-regions, encoding each sub-region separately. In the second stage, the Text-Encoder encodes the text dataset to obtain text features. It then calculates the similarity between the image sub-region features and encoded text features, selecting the text features with the highest similarity. Subsequently, the pre-trained Faster RCNN model extracts local image features. The model then splices together the text features, global image features, and local image features to fuse the multimodal information. In the final stage, the extracted features are fed into the Captioning model, which effectively fuses the different features using a novel global weighted attention pooling layer. The Captioning model then generates natural language image captions. The proposed model is evaluated on the MS-COCO dataset, Flickr 30K dataset, and BUUISE-Image dataset, using mainstream evaluation metrics. Experiments demonstrate significant improvements across all evaluation metrics on the public datasets and strong performance on the BUUISE-Image traffic scene dataset.

Personalized Intelligent Recommendation Model Construction Based on Online Learning Behavior Features and CNN

Dianqing Bao — 2024-03-22

The current intelligent recommendation models in online learning systems suffer from data sparsity and cold start problems. To address the data sparsity problem, a collaborative filtering recommendation algorithm model (SACM-CF) based on an automatic coding machine is proposed in the study. The model can extract the online learning behavior features of users and match these features with the learning resource features to improve the recommendation precision. For the cold-start problem, the study proposes a CBCNN model based on CNN, using the language model as the input of the model and the implicit factor as the output of the model. To avoid the problem of over-smoothing the implicit factor model, which affects the recommendation precision, an improved matrix decomposition method is proposed to constrain the output of the CNN and improve the model precision. The RMSE of SACM-CF is 0.844 and the MAE is 0.625. The MAE value of CBCNN is 0.72, the recall value is 0.65, the recommendation precision is 0.954 and the F1-score is 0.84. The metrics of SACM-CF and CBCNN are better than the existing state-of-the-art recommendation models. SACM-CF and CBCNN outperform the existing state-of-the-art intelligent recommendation models in all metrics. Therefore, the SACM-CF model and the CBCNN model can effectively improve the precision of the online learning system in recommending interesting learning resources to users, thus avoiding users' wasted learning time in searching and selecting learning resources and improving users' learning efficiency.

A Hybrid Strategy Guided Multi-Objective Artificial Physical Optimizer Algorithm

Bao Sun — 2024-03-22

Artificial physical optimizer (APO), as a new heuristic stochastic algorithm, is difficult to balance convergence and diversity when dealing with complex multi-objective problems. This paper introduces the advantages of R2 indicator and target space decomposition strategy, and constructs the candidate solution of external archive pruning technology selection based on APO algorithm. A hybrid strategy guided multi-objective artificial physical optimizer algorithm (HSGMOAPO) is proposed. Firstly, R2 indicator is used to select the candidate solutions that have great influence on the convergence of the whole algorithm. Secondly, the target space decomposition strategy is used to select the remaining solutions to improve the diversity of the algorithm. Finally, the restriction processing method is used to improve the ability to avoid local optimization. In order to verify the comprehensive ability of HSGMOAPO algorithm in solving multi-objective problems, five comparison algorithms were evaluated experimentally on standard test problems and practical problems. The results show that HSGMOAPO algorithm has good convergence and diversity in solving multi-objective problems, and has the potential to solve practical problems.

RWESA-GNNR: Fusing Random Walk Embedding and Sentiment Analysis for Graph Neural Network Recommendation

Junlin Gu — 2024-03-22

A graph neural network-based recommendation system treats the relationship between user items as a graph, and achieves deep feature mining by modelling the graph nodes. However, the complexity of the features of graph neural network-based recommendation systems brings poor interpretability and suffers from data sparsity problems. To address the above problems, a graph convolutional neural network recommendation model (RWESA-GNNR) based on random walk embedding combined with sentiment analysis is proposed. Firstly, a random walk-based matrix factorization is designed as the initial embedding. Secondly, the user and item nodes are modelled using a convolutional neural network with an injected attention mechanism. Then, sentiment analysis is performed on the review text, and attention mechanism is introduced to fuse text sentiment features and semantic features. Finally, node features and text features are aggregated to generate recommendation results. The experimental results show that our proposed algorithm outperforms traditional recommendation algorithms and other graph neural network-based recommendation algorithms in terms of recommendation results, with an improvement of about 2.43%-5.75%.

Point Cloud Completion Based on Nonlocal Neural Networks with Adaptive Sampling

Na Xing — 2024-03-22

Raw point clouds are usually sparse and incomplete, inevitably containing outliers or noise from 3D sensors. In this paper, an improved SA-Net based on an encoder-decoder structure is proposed to make it more robust in predicting complete point clouds. The encoder of the original SA-Net network is very sensitive to noise in the feature extraction process. Therefore, we use PointASNL as the encoder, which weights around the initial sampling points through the AS module (Adaptive Sampling Module) and adaptively adjusts the weight of the sampling points to effectively alleviate the bias effect of outliers. In order to fully mine the feature information of point clouds, it captures the neighborhood and long-distance dependencies of sampling points through the LNL module (Local-NonLocal Module), providing more accurate information for point cloud processing. Then, we use the encoder to extract local geometric features of the incomplete point cloud at different resolutions.Then, an attention mechanism is introduced to transfer the extracted features to a decoder. The decoder gradually refines the local features to achieve a more realistic effect. Experiments on the ShapeNet data set show that the improved point cloud completion network achieves the goal and reduces the average chamfer distance by 3.50% compared to SA-Net.

A Robust Adaptive Non-Singular Terminal Sliding Mode Control: Application to an Upper-Limb Exoskeleton with Disturbances and Uncertain Dynamics

Mouna Dali Hassen — 2024-03-22

This paper presents a new control strategy for uncertain upper-limb exoskeleton systems, which are known to have high nonlinearities, unmodeled dynamics, and uncertainties. The proposed technique is based on the terminal sliding mode control algorithm and its nonsingular design method and incorporates an adaptive control approach to estimate the upper bounds of the unknown system uncertainties, which helps to improve the accuracy of the control and reduce the effects of disturbances. The stability of the proposed control strategy is confirmed using Lyapunov theory, and its effectiveness is tested on a two-degrees-of-freedom upper-limb exoskeleton. The results demonstrate that the proposed control scheme provides robust, fast, and finite-time convergence as well as an effective control approach capable of dealing with the disturbances and uncertainties that such systems are prone to.

A Systematic Review and Meta-Analysis of Eye-Tracking Studies for Consumers’ Visual Attention in Online Shopping

Xin Li — 2024-03-22

As the marketing landscape continues to evolve, consumer preference remains a key driver of corporate profitability. Extensive research has shown that visual attention is a critical factor in consumer decision-making. However, a comprehensive meta-analysis of online shopping visual presentation has yet to be conducted. This paper applies various eye-tracking dependent variables to investigate consumer visual attention in relation to four common interface design factors: brand, endorser, product, and text. Generally, from the research it shown that product and brand havd positive effect, while text might be negative. It is worthy mention that we identified the subgroup analysis involving total time of fixation (SMD=-0.020, 95%CI: [-0.079,0.039], p=0.507), fixation count (SMD=-0.032, 95%CI: [-0.109,0.045], p=0.421) and time to first fixation (SMD=0.464, 95%CI: [0.346,0.582], p=0.000). In this paper, exposure time obviously impacted FC (Q-value=11.637, p=0.003) and TTFF (Q-value=10.316, p=0.006) in the reanalysis studies. Meanwhile consumer preference highly related to FC (Q=10.953, p=0.001) and TTFF (Q=6.540, p=0.011) were under concern. Studies contained 17 papers with a total of 1071 participants. The publication bias was within the reasonable rang and the heterogeneity mainly resulted in subgroup and moderator differences. Our study on systematic review and meta-analysis show that, to appropriately control the consumer visual attention attributes could be a good solution for increasing consumer preference in online shopping interaction experience. Furthermore, more controllable design factor and moderators related to visual attention should be concerned for neuromarketing progress. In the future, other measurements such as ERPs, FMRI, fINRs could be explored for making better consumer sentiment experience.

Tri-CLT: Learning Tri-Modal Representations with Contrastive Learning and Transformer for Multimodal Sentiment Recognition

Zhiyong Yang — 2024-03-22

Multimodal Sentiment Analysis (MSA) has become an essential area of research to achieve more accurate sentiment analysis by integrating multiple perceptual modalities such as text, vision, and audio. However, most previous studies failed to align the various modalities well and ignored the differences in semantic information, leading to inefficient fusion between modalities and generating redundant information. In order to solve the above problems, this paper proposes a transformer-based network model, Tri-CLT. Specifically, this paper designs Integrating Fusion Block to fuse modal features to enhance their semantic information and mitigate the secondary complexity of paired sequences in the transformer. Meanwhile, the cross-modal attention mechanism is utilized for complementary learning between modalities to enhance the model performance. In addition, contrastive learning is introduced to improve the model's representation of learning ability. Finally, this paper conducts experiments on CMU-MOSEI aligned and unaligned data, and the experimental results show that the proposed method outperforms the existing methods.

YOLOv5s-MEE: A YOLOv5-based Algorithm for Abnormal Behavior Detection in Central Control Room

Ping Yuan — 2024-03-22

Aiming to quickly and accurately detect abnormal behaviors of workers in central control rooms, such as playing mobile phone and sleeping, an abnormal behavior detection algorithm based on improved YOLOv5 is proposed. The technique uses SRGAN to reconstruct the input image to improve the resolution and enhance the detailed information. Then, the MnasNet is introduced to replace the backbone feature extraction network of the original YOLOv5, which could achieve the lightweight of the model. Moreover, the detection accuracy of the whole network is enhanced by adding the ECA-Net attention mechanism into the feature fusion network structure of YOLOv5 and modifying the loss function as EIOU. The experimental results in the custom dataset show that compared with the original YOLOv5 algorithm, the algorithm proposed in this paper improves the
detection speed to 75.50 frames/s under the condition of high detection accuracy, which meets the requirements of real-time detection. Meanwhile, compared with other mainstream behavior detection algorithms, this algorithm also shows better detection performance.

Learning Stabilization Control of Quadrotor in Near-Ground Setting Using Reinforcement Learning

Mantas Briliauskas — 2024-03-22

With the development of intelligent systems, the popularity of using micro aerial vehicles (MAV) increases significantly in the fields of rescue, photography, security, agriculture, and warfare. New modern solutions of machine learning like ChatGPT that are fine-tuned using reinforcement learning (RL) provides evidence of new trends in seeking general artificial intelligence. RL has already been proven to work as a flight controller for MAV performing better than Proportional Integral Derivative (PID)-based solutions. However, using negative Euclidean distance to the target point as the reward function is sufficient in obstacle-free spaces, e.g. in the air, but fails in special cases, e.g. when training near the ground. In this work, we address this issue by proposing a new reward function with early termination. It not only allows to successfully train Proximal Policy Optimization (PPO) algorithm to stabilize the quadrotor in the near-ground setting, but also achieves lower Euclidean distance error compared to the baseline setup.

An Efficient Deep Learningbased Intrusion Detection System for Internet of Things Networks with Hybrid Feature Reduction and Data Balancing Techniques

Hamdullah Karamollaoğlu — 2024-03-22

With the increasing use of Internet of Things (IoT) technologies, cyber-attacks on IoT devices are also increasing day by day. Detecting attacks on IoT networks before they cause any damage is crucial for ensuring the security of the devices on these networks. In this study, a novel Intrusion Detection System (IDS) was developed for IoT networks. The IoTID20 and BoT-IoT datasets were utilized during the training phase and performance testing of the proposed IDS. A hybrid method combining the Principal Component Analysis (PCA) and the Bat Optimization (BAT) algorithm was proposed for dimensionality reduction on the datasets. The Synthetic Minority Over-Sampling
Technique (SMOTE) was used to address the problem of data imbalance in the classes of the datasets. The Convolutional Neural Networks (CNN) model, a deep learning method, was employed for attack classification. The proposed IDS achieved an accuracy rate of 99.97% for the IoTID20 dataset and 99.98% for the BoT-IoT dataset in attack classification. Furthermore, detailed analyses were conducted to determine the effects of the dimensionality reduction and data balancing models on the classification performance of the proposed IDS.

Optimizing Parkinson's Disease Diagnosis with Multimodal Data Fusion Techniques

C.M.T. Karthigeyan — 2024-03-22

Parkinson's disease (PD) is a central nervous system neurodegenerative illness. Its symptoms include poor motor skills, speech, cognition, and memory. The condition is incurable, although evidence shows that early identification and therapy reduce symptoms. A lack of medical facilities and personnel hinders PD identification. PD is a common chronic degenerative neurological dyskinesia that threatens the elderly. Multi-modal data fusion may reveal more about PD pathophysiology. This study aims to contribute to the evaluation of PD by introducing a novel multimodal deep-learning technique for distinguishing individuals with PD from those without PD. This study utilizes resting functional magnetic resonance imaging (rfMRI) and gene data obtained from the Parkinson's Progression Markers Initiative (PPMI) and Alzheimer's Disease Neuroimaging Initiative (ADNI) databases. The primary objective is to predict the specific pathological brain regions and identify the risk genes associated with PD. The authors want to learn more about the genetic components and underlying procedures by analyzing these datasets. Contributing to the development and progression of PD. In this study, we present our findings that demonstrate the superior recital of our proposed multimodal method compared to both unimodal approaches and other existing multimodal methods. Our evaluation is based on an extensive dataset consisting of real patients. Specifically, our proposed method stacked deep learning classifiers (SDLC) achieves an impressive F1-score of 0.99 and an accuracy of 99.4%, surpassing the performance of both unimodal approaches and other multimodal methods. These results highlight the efficiency and potential of our method in enhancing the accuracy and reliability of patient data analysis. In this study, we demonstrate that our proposed method consistently surpasses alternative approaches in terms of performance, as indicated by a higher average increase in F1-score. This finding highlights the advantage of training on multiple modalities, even when a particular modality is absent during inference.

A Multi-level Surrogate-assisted Algorithm for Expensive Optimization Problems

Liang Hu — 2024-03-22

With the development of computer science, more and more complex problems rely on the help of computers for solving. When facing the parameter optimization problem of complex models, traditional intelligent optimization algorithms often require multiple iterations on the target problem. It can bring unacceptable costs and resource costs in dealing with these complex problems. In order to solve the parameter optimization of complex problems, in this paper we propose a multi-level surrogate-assisted optimization algorithm (MLSAO). By constructing surrogate models at different levels, the algorithm effectively explores the parameter space, avoiding local optima and enhancing optimization efficiency. The method combines two optimization algorithms, differential evolution (DE) and Downhill simplex method. DE is focused on global level surrogate model optimization. Downhill simplex is concentrated on local level surrogate model update. Random forest and inverse distance weighting (IDW) are constructed for global and local level surrogate model respectively. These methods leverage their respective advantages at different stages of the algorithm. The MLSAO algorithm is evaluated against other state-of-the-art approaches using benchmark functions of varying dimensions. Comprehensive results from the comparisons showcase the superior performance of the MLSAO algorithm in addressing expensive optimization problems. Moreover, we implement the MLSAO algorithm for tuning precipitation parameters in the Community Earth System Model (CESM). The outcomes reveal its effective enhancement of CESM's simulation accuracy for precipitation in the North Indian Ocean and the North Pacific region. These experiments demonstrate that MLSAO can better address parameter optimization problems under complex conditions.

Dual Attention Aware Octave Convolution Network for Early-Stage Alzheimer's Disease Detection

Banupriya Rangaraju — 2024-03-22

Some of the most fundamental human capabilities, including thought, speech, and movement, may be lost due to brain illnesses. The most prevalent form of dementia, Alzheimer's disease (AD), is caused by a steady decline in brain function and is now incurable. Despite the challenges associated with making a conclusive diagnosis of AD, the field has generally shifted toward making diagnoses justified by patient records and neurological analysis, such as MRI. Reports of studies utilizing machine learning for AD identification have increased in recent years. In this publication, we report the results of our most recent research. It details a deep learning-based, 3D brain MRI-based method for automated AD detection. As a result, deep learning models have become increasingly popular in recent years for analyzing medical images. To aid in detecting Alzheimer's disease at an initial phase, we suggest a dual attention-aware Octave convolution-based deep learning network (DACN). The three main parts of DACN are as follows: First, we use Patch Convolutional Neural Network (PCNN) to identify discriminative features within each MRI patch while simultaneously boosting the features of abnormally altered micro-structures in the brain; second, we use an Octave convolution to minimize the spatial redundancy and widen the field of perception of the brain's structure; and third, we use a dual attention aware convolution classifier to dissect the resulting depiction further. An outstanding test accuracy of 99.87% is reached for categorizing dementia phases by employing the suggested method in experiments on a publically available ADNI (Alzheimer's Disease Neuroimaging Initiative) dataset. The proposed model was more effective, efficient, and reliable than the state-of-the-art models through our comparisons.