Abdulazeez , Adnan Mohsin
Unknown Affiliation

Published : 4 Documents Claim Missing Document
Claim Missing Document
Check
Articles

Found 4 Documents
Search

Facial Beauty Prediction Based on Deep Learning: A Review Arabo, Wahab; Abdulazeez , Adnan Mohsin
The Indonesian Journal of Computer Science Vol. 13 No. 1 (2024): The Indonesian Journal of Computer Science (IJCS)
Publisher : AI Society & STMIK Indonesia

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.33022/ijcs.v13i1.3743

Abstract

This review delves into Facial Beauty Prediction (FBP) using deep learning, specifically focusing on convolutional neural networks (CNNs). It synthesizes recent advancements in the field, examining diverse methodologies and key datasets like SCUT-FBP and SCUT-FBP5500. The review identifies trends in FBP research, including the evolution of deep learning models and the challenges of dataset biases and cultural specificity. The paper concludes by emphasizing the need for more inclusive and balanced datasets and suggests future research directions to enhance model fairness and address ethical implications.
A Comparative Study of Multi-Class Classification Based on Imbalanced Data: A Review Abdulkareem, Rojan; Abdulazeez , Adnan Mohsin
The Indonesian Journal of Computer Science Vol. 14 No. 5 (2025): The Indonesian Journal of Computer Science
Publisher : AI Society & STMIK Indonesia

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.33022/ijcs.v14i5.5020

Abstract

Classification of unbalanced multiclass datasets is still a major challenge in machine learning in many fields of applications, including medical diagnostics, fraud detection, and picture classification, where minority classes are the most crucial, but at the same time under-represented. Classical classification algorithms designed for balanced data tend to overfit the majority classes deeming a large number of minority classes misclassified and, as a result, compromising the model's performance. This review covers the main state-of-the-art techniques for class imbalance problems including under-sampling and over-sampling techniques, ensemble approaches, cost-sensitive learning, and producing synthetic data via SMOTE (synthetic minority oversampling technique). Recently, GANs (Generative Adversarial Networks) have also been employed to generate synthetic data, specifically valuable for complex datasets where realistic data augmentation is needed. Each of these techniques is analyzed in terms of their capability of dealing with imbalanced data through conventional metrics such as accuracy and specific metrics for imbalanced datasets such as F1-score, G-mean, and others. Recent advancements, such as hybrid approaches and learning from deep learning models are also discussed as viable solutions given the complexities associated with big data (high dimensional and large) and their corresponding models. Such comparative analysis should facilitate the construction of more robust models that handle complex data in modern applications.
A Facial Expression Prediction Based on Pre-Trained ResNet50 and SVM Ihsan, Rasheed; Abdulazeez , Adnan Mohsin
The Indonesian Journal of Computer Science Vol. 14 No. 5 (2025): The Indonesian Journal of Computer Science
Publisher : AI Society & STMIK Indonesia

Show Abstract | Download Original | Original Source | Check in Google Scholar

Abstract

Facial expression prediction has become a vital area in computer vision, with applications spanning healthcare, security, and human-computer interaction. This study proposes a robust system for binary facial expression prediction using a combination of classical computer vision techniques and deep learning. The system employs the Haar Cascade algorithm for face detection and ResNet50, a 50-layer deep residual network, for feature extraction. Support Vector Machines (SVM) with a radial basis function kernel are used for classification. Using the 4,000 tagged images from the GENKI dataset, preprocessing and data augmentation improved the model's capacity for generalization. Experimental results demonstrate the system’s effectiveness, achieving a test accuracy of 94.65%. The robust integration of classical and modern techniques ensures computational efficiency while maintaining high performance. For real-world applications, this method provides a scalable solution that tackles issues including lighting fluctuation, position, and expression variation.
Hybrid Transfer Learning Model for Facial Attractiveness Prediction Hawar Bahzad Ahmad; Abdulazeez , Adnan Mohsin
The Indonesian Journal of Computer Science Vol. 14 No. 6 (2025): The Indonesian Journal of Computer Science
Publisher : AI Society & STMIK Indonesia

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.33022/ijcs.v14i6.5018

Abstract

Prediction of facial attractiveness greatly depends on the subjective terminology applied according to the diverse cultural, social and psychological considerations. This task is important for applications in many fields, such as aesthetics, entertainment, wardrobe recommendations, etc., and requires accurate and robust models. Current methods predominantly adopt a single model, which is unable to learn the diverse attributes that can influence the quality of facial beauty. In order to overcome these challenges, this study proposes a hybrid transfer learning framework for feature extraction and prediction that combines ResNet50 and InceptionV3. In this methodology, Multi-task Cascaded Convolutional Networks (MTCNN) is used for accurate face detection and preprocessing, then features extraction is done using pretrained ResNet50 and InceptionV3 architectures. The features extracted are then normalized and fused together and passed through a dense classification layer with application of dropouts and regularization in order to make the model robust. The CelebA dataset was used to train the model, utilizing class weights to account for imbalanced data and callbacks to optimize performance. Test accuracy and F1 Score of the proposed model is found to be 83.58% and 0.8384 respectively, which shows good generalization on unseen data. The validation frames the performance of the hybrid framework which leverages the complementary strengths of multiple CNNs, and thus provides robust performance.