Author ORCID Identifier
0009-0005-9422-7015
Biosketch
I am Ajantha V, a Computer Science and Engineering professional with a strong academic and research background in Artificial Intelligence, Machine Learning, Deep Learning, and Computer Vision. I completed my B.Tech. in Information Technology and M.E. in Computer Science and Engineering, and I have completed my PhD in Computer Science and Engineering at SASTRA Deemed to be University.
I have three years of research experience as a Junior Research Fellow at SASTRA under the MHI Project, where I worked on deep learning and real-time computer vision, particularly automated defect detection and inspection in pharmaceutical packaging. My research focuses on developing intelligent, efficient, and industry-relevant deep learning solutions.
I also have teaching experience as an Assistant Professor in Computer Science and Engineering, where I handled subjects such as Machine Learning and Python Programming, conducted laboratory sessions, and guided students in academic projects.
My research contributions include SCI/SCIE-indexed journal publications, Scopus-indexed conference publications, a book chapter, and a filed Indian patent. My areas of interest include Deep Learning, Machine Learning, Computer Vision, Cloud Computing, and Networking. I am passionate about combining teaching, research, and practical applications of emerging technologies.
Date of Award
1-9-2026
Document Type
Thesis
School
School of Computing
Programme
Ph.D.-Doctoral of Philosophy
First Advisor
Dr.K.Joseph Abraham Sundar
Keywords
Defect Detection, Pharma 4.0, Deep Learning, Computer Vision, CNN
Abstract
Defect detection plays an important role in maintaining product quality and safety. In the pharmaceutical industry, oral dosage forms, particularly tablets, are among the most commonly used methods of drug delivery. However, visual defects in blister-packed tablets can affect drug quality, therapeutic effectiveness and patient safety. Therefore, accurate and real-time detection of defects in pharmaceutical blister packages is essential. Traditional inspection methods mainly depend on manual checking, which is time-consuming, labour-intensive and prone to human errors. Automated inspection also faces several challenges, including high computational requirements, reflections from blister surfaces, limited training data, variations in defect shapes and sizes, and difficulties in tracking and counting defects without duplication or omission. This research focuses on developing artificial intelligence-based methods for detecting, tracking and counting visual defects in pharmaceutical blister packages.
To address these challenges, four complementary deep learning frameworks are developed, each focusing on a specific aspect of automated defect inspection. The first framework, CBS-YOLOv8, is designed to improve real-time defect detection by integrating Coordinate Attention for extracting important spatial and channel features, SimSPPF for reducing computational complexity, and Bidirectional Feature Pyramid Network (BiFPN) for effective multi-scale feature extraction. These modifications improve detection accuracy while maintaining the processing speed required for continuous industrial inspection. The second framework, PharmaNet Deep, addresses the challenges of limited training data and reflections from blister surfaces. It incorporates Defect-Guided Dynamic Feature Fusion (DGDFF) to highlight important defect features, Adaptive Deep Chain (ADC) to reconstruct useful information from partially hidden features, and an Uncertainty-aware Detection Head (UDH) to improve the reliability of detection results. The framework achieves high detection accuracy with a relatively low computational requirement, using approximately 15.1 million parameters.
The third framework, Multi-Scale Pharmaceutical Defect Tracker with Adaptive Confidence and Counting (MS-PDTAC), extends defect detection to continuous video-based inspection. It combines YOLO-based object detection, Kalman filtering for motion prediction, the Hungarian algorithm for object matching and spatio-temporal attention for improved tracking. A multi-method line-crossing strategy is also introduced to count defects accurately and avoid repeated or missed counts while tablets move along the conveyor belt. The fourth framework, LightDefectNet-18, focuses on developing a lightweight and generalised deep learning model for different industrial and agricultural applications. It is integrated with Faster R-CNN and uses dual residual blocks, progressive channel scaling and Feature Pyramid Network (FPN) integration to improve feature extraction at multiple scales. With only 4.9 million parameters, the model provides an efficient solution and demonstrates its ability to perform across six different application domains beyond pharmaceutical packaging.
The performance of all four frameworks is evaluated through extensive experiments. CBS-YOLOv8 achieves 97.4% mean Average Precision (mAP) on the custom pharmaceutical blister dataset while maintaining real-time detection speed. PharmaNet Deep achieves 99.4% mAP on the same dataset and 97.2% mAP on the publicly available MVTec AD benchmark, demonstrating its effectiveness even with limited training data. MS-PDTAC achieves 100% counting accuracy with zero absolute counting error across seven defect categories. LightDefectNet-18 achieves an overall mAP of 98.68% across six different application domains, demonstrating its ability to generalise across diverse industrial and agricultural datasets.
Overall, this research presents four complementary deep learning frameworks for accurate, efficient and reliable visual defect detection, tracking and counting. The proposed methods address important challenges in automated pharmaceutical inspection while also demonstrating their applicability to other industrial and agricultural sectors. The research contributes towards the development of intelligent, scalable and efficient quality inspection systems in line with Pharma 4.0, supporting the advancement of automated quality assurance and improving the reliability of modern manufacturing processes.
Recommended Citation
V, Ajantha Ms, "An Intelligent Computer Vision Framework for Real-Time Defect Detection, Tracking, and Counting in Pharmaceutical Blister Packaging" (2026). Theses and Dissertations. 230.
https://knowledgeconnect.sastra.edu/theses/230