Advanced Patterned Fabric Defect Detection and Calculating the Defect Size using Explainable AI
This project presents an AI-based approach to detect and measure defects in multi-patterned fabrics using computer vision, deep learning, and Explainable AI (XAI). Traditional fabric inspection processes rely heavily on manual labour, which is often slow, inconsistent, and prone to error—especially when dealing with complex or coloured patterns. Furthermore, manual inspection typically lacks precise defect sizing, which is critical for fabric quality evaluation based on standards like the 4-point system.
To address these limitations, a deep learning model based on the Xception architecture was trained on a publicly available patterned fabric dataset to classify fabric patterns and detect common defect types such as holes, stains, and multiple defects. Grad-CAM, a popular XAI method, was used to generate visual explanations of the model’s predictions, improving transparency and user trust. These heatmaps also enabled the localisation and measurement of defects, which were further converted from pixel dimensions into real-world units using camera parameters including optical working distance and focal length.
The model achieved 0.91 training accuracy and 0.88 accuracy on both validation and test sets for defect detection, with precision, recall, and F1-score also at 0.88. Pattern classification reached 0.98 test accuracy. This integrated system not only automates defect detection and sizing but also improves interpretability, making it a reliable and scalable tool for quality control in modern textile manufacturing.
Bias Lens: Systemic Bias Detection with Explainable Analysis
Artificial Intelligence, particularly Natural Language Processing, often inherits and amplifies societal biases, undermining fairness and trust. Existing bias detection tools frequently lack the granularity and transparency needed for effective mitigation, often operating as ‘black boxes’. Bias Lens is a pioneering framework designed to address this critical challenge. It leverages a fine-tuned BERT-based model for multi-label token classification, enabling the identification of six distinct, potentially overlapping, bias types (Generalization, Unfairness, Stereotype, Assumption, Exclusion, Framing) directly at the word level. A key innovation lies in its integration of advanced Explainable AI (XAI) techniques, including a novel enhanced Integrated Gap Gradients (IG2). This provides unprecedented clarity, attributing bias to specific tokens and explaining why content is flagged. Achieving 93% accuracy and significantly faster inference (~8x) than baseline research models, Bias Lens translates complex analyses into actionable insights through an interactive visualization dashboard, accessible to diverse users. By delivering deep, interpretable bias detection and contributing open-source resources (validated dataset, model, XAI library), Bias Lens empowers stakeholders to proactively build fairer, more transparent, and ethically sound AI, fostering responsible technological advancement.
Floro-X
The floriculture industry faces a major issue due to the extremely perishable characteristics of its products, covering inventory uncertainty due to demand variability across consumers and seasonality. Seasonal fluctuations have the impact of leading to enormous wastage of flowers, monetary loss, and inefficiency in handling stocks. The traditional method of inventory control, which is still in use in small and medium-scale florist enterprises lacks the tools, technologies and prediction capabilities required to take necessary measures to address such issues. As a solution to these constraints, this project introduces Floro-X, a forecasting-based inventory management system for the floriculture industry.
Floro-X is a web-based software solution that offers a demand prediction solution to florist managers in order to align inventory levels with real-world consumer demand. Using Facebook Prophet, a well-tested time series forecasting algorithm, the machine learning model is trained from historical sales trends and seasonal patterns for five types of flowers. The prediction algorithm delivers forecast accuracies above 85% for almost all flowers enabling users to make better inventory and sales planning decisions as to avoid overstocking or running out of stock. System features a user-friendly interface accessible via desktop, with fundamental modules of stock management, demand planning, and user profile handling. Florist managers are able to update the inventory details, place future demand forecast orders, and examine previously calculated outcomes, all from a single, user-friendly console.
While the initial version of Floro-X runs with fixed data sets which was used for training, future enhancements will involve the integration with Point-of-Sale (POS) systems to enable real-time tracking of sales and retraining of the model based on new sales data in an automated process, inclusion of real-time weather as a forecasting variable, and expanding prediction capability to include more varieties of flowers. Lastly, Floro-X addresses a major issue in the flower supply chain by introducing digital transformation to a traditionally manual and labour-intensive industry, helping florists make smarter, on-time, data driven decisions. At its core, the adoption of Floro-X facilitates a more profitable, robust, and sustainable floral system, a platform for future innovation and growth within the floriculture sector across the globe.
Enhancing Personalized, Contextual, and Temporal Recommendations Using Multi-Task Learning: A MovieLens Case Study
Traditional recommender systems struggle to adapt to users’ changing preferences due to their reliance on single-task learning. This research proposes a multi-task learning (MTL) approach that jointly addresses personalized rating prediction, contextual tag recommendation, and temporal preference modeling. By using shared embeddings and task-specific branches, the model better captures dynamic user behavior and context, improving recommendation accuracy and relevance.
FED-Ensemble: A Federated Learning Ensemble Architecture with Generative Models for Poisoning Attack Mitigation
Federated Learning (FL) endorses promising avenues for organizations, enabling collaborative model training across a distributed client network, exempting clients from relaying raw data, thus preserving privacy and lowering communication overhead. Despite these benefits, the decentralized nature of FL systems remains highly susceptible to adversarial threats, particularly Poisoning Attacks (PA) from the network’s edge. Countermeasures against PA are often tailored to specific client behaviors, yet the global model’s integrity remains at risk due to evolving malicious strategies.
The author aims to bridge this gap by proposing a feature representation learner using a novel unsupervised detection framework. It leverages an Ensemble of Deep Generative Models (EDGMs) to assign suspicious scores to model updates, eliminating irregularities and aggregating only legitimate contributions. The objective is to deflect anomalies through continuous representation learning and provide a practical solution grounded in theoretical models.
The system underwent preliminary testing under limited experimental configurations, in line with available resources. Results showed promising detection accuracy—around 90% on MNIST and 70–80% on CIFAR-10—while mitigating poisoning even when 20–40% of clients were compromised. Future work includes improving client data diversity, boosting detection efficiency, and enhancing system scalability.
AKURU: Addressing Ad hoc Back-Transliteration with Word Sense Disambiguation in Romanized Sinhala Through a Context-Aware Approach
The growing use of Romanized Sinhala in digital communication platforms brings significant challenges to natural language processing tasks, particularly in backward transliterations which is the process of converting Romanized Sinhala text into the native Sinhala script. Since Romanized Sinhala is an informal way to represent native Sinhala, it lacks standardized spelling conventions leading to ad-hoc typing variations in which users frequently omit vowels, apply inconsistent phonetic spellings and use alternative consonants. This inconsistency exacerbates the problem of lexical ambiguity making it difficult to interpret meanings from Romanized Sinhala text. Existing back-transliteration systems struggle with these ad-hoc typing variations and word sense disambiguation, leading to significant accuracy loss.
To address these challenges in existing back-transliteration systems, this research introduces a novel context-aware hybrid approach that combines an ad-hoc transliteration dictionary and rule-based approach with BERT-based language model trained on native Sinhala text.
The proposed system was evaluated for backward transliteration using Sinhala BERT and its fine-tuned variant, achieving BLEU scores of around 0.91 with remarkably low Word Error Rate and Character Error Rate, approximately 0.09 and 0.02 respectively. Additionally, the first ever Word Sense Disambiguation (WSD) dataset for Romanized Sinhala is introduced as a part of this research. The proposed transliterator achieved an overall F1 score of approximately 0.94 highlighting the effectiveness of the proposed approach in handling ambiguous words in Romanized Sinhala.
StackTraceXAI
An explainable hybrid fraud detection system combining a Variational Autoencoder (VAE) and XGBoost.
The VAE learns the underlying patterns of non-fraudulent ERP financial transactions and generates reconstruction errors and latent features, which are then used to enhance the accuracy of an XGBoost classifier. The system integrates SHAP and LIME based explainable AI to provide transparent and interpretable fraud predictions.
Cholecheck : Explainable AI-Based Multi-Class Classification of Gallbladder Diseases Using Ultrasound Images
Gallbladder diseases affect a significant portion of the global popula
tion, ranging from gallstones and cholecystitis to carcinoma. Accurate diagnosis
is critical but often delayed due to overlapping symptoms and limitations in ul
trasound imaging interpretation. This study proposes an advanced multi-class
classification system leveraging deep learning techniques to identify nine distinct
gallbladder disease types using ultrasound images. The methodology integrates
preprocessing techniques such as CLAHE and active contour segmentation to
enhance image quality and isolate regions of interest. A hybrid ensemble model
combining VGG16, ResNet152, and a custom CNN achieved superior perfor
mance with 99.8% accuracy and a Cohen Kappa score of 99.8%. Transfer learn
ing, feature fusion, and ensemble strategies were employed to improve robust
ness and generalization. Explainable AI (XAI) techniques like Grad-CAM and
LIME were incorporated to provide interpretable visualizations of the model’s
predictions, aiding clinical decision-making. The system was trained on the UI
DataGB dataset containing 10,692 annotated ultrasound images, ensuring high
reliability across diverse gallbladder conditions. Comparative benchmarking
demonstrates that the proposed model outperforms existing systems in accuracy
and classification depth. This research contributes a scalable, interpretable AI
driven diagnostic tool that enhances early detection and management of gallblad
der diseases while addressing challenges in medical imaging variability.
AIRA – Generalized detection of AI-generated images leveraging Explainable AI
Artificially generated images have rapidly risen in both quality and quantity, largely driven by their popularity across domains due to their ease of access and use. This poses threats of misinformation and lack of trust in artificially generated images that pass as real, particularly in news and social media. Current research focuses on the development of methods to enhance detection, improving generalization, or on the addition of interpretability factors using explainable artificial intelligence (XAI). This project addresses the need for a detection system that allows for efficient prediction of “AI-generated” images across a variety of image generators and of various content types over “Real” images, whilst providing insights and explanations for the prediction.
In this research, a novel approach is proposed for the detection of AI-generated images while integrating Explainable Artificial Intelligence techniques to facilitate interpretability. The proposed solution implementation was developed using a Convolutional Neural Network (CNN) classifier built on top of the DenseNet121 architecture, which leveraged pre-trained weights from ImageNet, allowing for efficient feature extraction. The model was trained on a curated dataset consisting of images from five AI-image generators, which included Midjourney, Stable diffusion, Dall-E, ProGAN, and other images from State-of-The-Art image generators.
The proposed solution was evaluated using a range of data science metrics, including the F1 score, precision, accuracy and recall on an unseen portion of the dataset. The initial results prove promising, achieving on average an accuracy of 93% while maintaining an F1-score of 93% and loss of 0.22.
MMAD : Multi-model Adversarial Defense for Medical Images
Adversarial attacks on medical imaging pose a critical threat to AI-guided diagnosis, as these invisible perturbations can cause catastrophic misclassifications. This research presents the Multi-model Adversarial Defense (MMAD) system, introducing novel algorithmic innovations for medical imaging protection, specifically targeting Magnetic Resonance Imaging (MRI).
The proposed solution features a groundbreaking hybrid classifier integrating Vision Transformer, Convolutional Neural Network, and Spiking Neural Network through a novel attention-based fusion algorithm: α = Softmax(W_attn · [f_vit, f_cnn, f_snn]), where fused features f_fused = Σ α_i · f_i dynamically weight each model’s contribution. This architecture incorporates auxiliary classifiers with weighted focal loss, achieving remarkable benchmarking results on MNIST dataset: 98% accuracy for FGSM, 96% for BIM, 90.4% for PGD, and 100% for clean images at epsilon=0.05.
The two-phase purification framework introduces an innovative multi-branch output strategy: Purified Image = Main + 0.3 × Detail + 0.15 × Edge, utilizing a U-Net generator with dual attention mechanisms (channel+spatial) and self-attention at the bottleneck. Phase-two refinement employs a PatchGAN discriminator with spectral normalization, achieving peak PSNR of 34.61dB and SSIM of 0.9551.
Unlike existing solutions, MMAD eliminates the accuracy-security trade-off while providing generalized defense against multiple white-box attacks (FGSM, BIM, PGD) at low perturbation intensities (ε=0.01-0.05). The medical-first algorithmic design preserves anatomical structures and diagnostic features, establishing new benchmarks for adversarial defense in healthcare AI systems, ultimately enhancing patient safety and diagnostic reliability.