International Journal of Intelligent Systems Technologies and Applications (5 papers in press)

Regular Issues

• Toward Arabic Social Networks unmasking Toxicity using machine learning and deep learning models  
  by Anis Mezghani Mezghani, Mohamed Elleuch, Saloua Gasmi, Mongi Kherallah 
  Abstract: With the rapid expansion of social media usage, the issue of racism has gained momentum, leading to an increased prevalence of racist discussions. Despite the efforts made by international organisations and prominent social media platforms like Twitter and Facebook to combat racism, it persists as a real-world problem. Consequently, a multitude of researchers are now directing their attention towards the detection of hate speech and racism on social media, particularly in the context of the Arabic language. We therefore propose the creation of an intelligent system for detecting and classifying toxic comments in Tunisian dialect using deep learning and machine learning models. Specifically, we use LR, linear SVC, SVM and BLSTM. This analysis has been performed using the potent technique for text analysis named NLP. The best performances were provided by combining BLSTM and SVM models with an accuracy of 99.83% for binary classification and 98.42% for ternary classification.
  Keywords: racism; Arabic content; social media; natural language processing; NLP; machine learning; deep learning.
  DOI: 10.1504/IJISTA.2024.10061713
   
  
• Yolov5-based Convolutional Feature Attention Neural Network for Plant Disease Classification  
  by Jameer Gulab Kotwal, Ramgopal Kashyap, Pathan Mohd. Shafi 
  Abstract: This article employs pre-processing feature extraction and classification to identify plant diseases. Pre-processing involves rescaling, contrast enhancement and filtering based on a bilateral new fast filter (BNFF). Rescaling and contrast enhancement improve contrast, and a bilateral new quick filter filters the image without degrading the image quality. Feature extraction and classification are performed using a hybrid classification network, the Yolov5-based convolutional feature attention network (Yolov5-CFAN). YOLO V5 determines the diseased portion, and CFAN is used to perform feature extraction from the detected portion and image classification. The experimental results section compares existing models to the proposed model using accuracy, precision, recall, specificity, and F1-score. The proposed model attained an accuracy of 99.55%, a precision of 97.55%, a specificity of 99.99% and a sensitivity of 97.5%. The research also trained the suggested model to recognise early, healthy, and late blight.
  Keywords: contrast features; spatial domain; denoising; overfitting; attention layer; feature map.
  DOI: 10.1504/IJISTA.2024.10062157
   
  
• A Novel Framework of Video Condensation and Video Retrieval Process using Hybrid Meta-Heuristic Development with Yolo-based Anomaly Detection  
  by Suhan Das, Santhosh Kumar G 
  Abstract: Visual surveillance systems have presently got the attention of various researchers. Anomaly detection is considered as a challenging g issue that reduces the accuracy rate in the video retrieval system with the video surveillance data. To resolve this, a novel method is proposed for the video condensation and retrieval model. The extracted frames are given as input to the YoloV5 model, where the anomalies objects are detected. Here, the condensed video is retrieved using anomaly aware EOHLO-DGRU-MSF (AA-EOHLO-DGRU-MSF). Then, the parameters in DenseNet and GRU are optimised by using equilibrium optimiser-assisted hybrid leader optimisation (EOHLO) algorithm for attaining optimal results. Finally, the multi-similarity function (MSF) between the features in the database and query video is considered. Finally, the performance is evaluated and measured with diverse metrics. Contrary to other approaches, the proposed work outperforms the detection of video and retrieval of video.
  Keywords: video condensation; video retrieval; anomaly detection; DenseNet; gated recurrent unit; equilibrium optimiser assisted hybrid leader optimisation; YOLOv5 model; multi-similarity function; MSF.
  DOI: 10.1504/IJISTA.2024.10062805
   
  
• Adaptive semi-supervised facial expression recognition method based on improved ResNet50  
  by Zeqiang Lin, Siwen Wang 
  Abstract: This study addresses the limitations of single convolutional neural networks in deep learning, which struggle with inadequate extraction of features from imbalanced expression labels and exhibit recognition errors when subjected to disturbances. The proposed solution introduces an adaptive semi-supervised facial expression recognition model. This model adeptly extracts expression features from unbalanced datasets, mitigates overfitting, and thus enhances overall expression recognition accuracy. By incorporating a self-attention mechanism, optimising convolutional kernels, and introducing replacement activation functions within the ResNet50 network, both computational efficiency and feature extraction are significantly improved. Moreover, the application of the adaptive semi-supervised method within training refines the accuracy of the model and prevents overfitting, thereby bolstering its robustness. Experimental findings indicate that the adaptive semi-supervised network, based on the enhanced ResNet50, achieves recognition rates of 73% and 99.57% on the FER2013 and JAFFE facial expression datasets, respectively. Comparative analysis with traditional single convolutional neural networks like ResNet18, VGG16, and VGG19, as well as optimised networks like IL-CNN, reveals an overall accuracy improvement of 27% and 14%, respectively.
  Keywords: semi-supervised learning; expression recognition; residual networks; adaptive methods; convolution kernel.
  DOI: 10.1504/IJISTA.2024.10062993
   
  
• Deep-Reinforcement Learning aided Dynamic Parameter Identification of Multi-Joints Manipulator  
  by Zhuoran Bi, Wenlong Zhao, Yichao Huang, Haoran Zhou, Li Qingdu 
  Abstract: To obtain more accurate dynamics equation parameters, this paper proposed a deep reinforcement learning (DRL) method for parameter identification. After using the least square (LS) method to identify the base parameters, we establish a training strategy where the friction coefficient serves as the DRL action. This strategy controls both the source and target manipulators, employing the concept of imitation learning. After using our strategy, the parameters of the target manipulator tend to converge to those of the source manipulator. In the experiment, we perform parameter identification of a 7-degree-of-freedom (DOF) manipulator in a real environment, and then identify friction coefficient for each joint based on the MuJoCo environment to theoretically validate the parameter identification using DRL. The identification results demonstrated that in a simulation environment, the use of DRL outperforms the traditional LS method, resulting in improved accuracy.
  Keywords: deep reinforcement learning; parameter identification; soft actor critic; joint friction.
  DOI: 10.1504/IJISTA.2024.10064051