\section*{Inspiration} Missed changes in manufacturing can lead to big risks and wasted resources. We wanted a practical tool that detects not just single anomalies, but shows how defects evolve across an image sequence.
\section*{What it does} \textbf{काल परिवर्तन चक्षु} uses multiple change detection methods (SSIM, histogram, edge, HOG features) and tracks anomaly evolution over time in MVTec AD industrial images. It predicts whether a detected defect is escalating or resolving and provides a clear severity score.
\section*{How we built it} \begin{itemize} \item Built in Python using scikit-image, OpenCV, SciPy, Matplotlib, and NumPy. \item Combined structural similarity (SSIM), texture/histogram, edge detection, and HOG features for robust anomaly detection. \item Developed a temporal evolution analyzer to track anomaly scores and trends across sequences. \item Visualized results using 6-panel plots, mapped scores, and change masks. \item Tested entirely on MVTec AD dataset for defect evolution. \end{itemize} Mathematical scoring foundation: [ A_{score} = 0.5 \cdot (1 - \text{SSIM}) + 0.3 \cdot \frac{D_{hist}}{1000} + 0.2 \cdot \frac{D_{HOG}}{50} ]
\section*{Challenges we ran into} \begin{itemize} \item Aligning normal and test images to minimize false change detections. \item Managing performance in Colab to process large, high-resolution image sequences. \item Setting thresholds for anomaly scoring across diverse defect types. \end{itemize}
\section*{Accomplishments that we're proud of} \begin{itemize} \item Ensemble detection outperformed single-method approaches for manufacturing images. \item Successfully tracked defect growth and severity using timeline analysis. \item Created visuals that clearly mark changes, severity, and evolution for MVTec AD examples. \end{itemize}
\section*{What we learned} \begin{itemize} \item Combining multiple image features detects a wider range of anomalies. \item Temporal scoring provides actionable insights for real-world inspections. \item Colab is a practical platform for large dataset analysis and rapid prototyping. \end{itemize}
\section*{What's next for काल परिवर्तन चक्षु} \begin{itemize} \item Generalize evolution analysis to other industrial domains. \item Expand anomaly scoring for multi-class defect prediction. \item Publish code and documentation for broader open-source use in manufacturing. \end{itemize}
Project Story: काल परिवर्तन चक्षु – Temporal Change Evolution Engine
Inspiration
Missed changes in manufacturing can lead to big risks and wasted resources. We wanted a practical tool that detects not just single anomalies, but shows how defects evolve across an image sequence—helping industrial and engineering teams address problems before they become catastrophic.
What it does
काल परिवर्तन चक्षु tracks how defects or structural changes evolve in visual inspection settings. Using a range of image-processing techniques (SSIM, histogram, edge, and HOG features), it analyzes MVTec AD industrial image sets to detect, score, and visualize temporal defect evolution. For every input sequence, the tool predicts whether a detected defect is escalating or resolving, and quantifies severity over time.
How we built it
- Engineered in Python, leveraging scikit-image, OpenCV, SciPy, Matplotlib, and NumPy for fast prototyping.
- Combined structural similarity (SSIM), texture/histogram difference, edge comparison, and HOG features for robust, diverse anomaly detection in each frame.
- Designed a temporal evolution analyzer to keep track of anomaly scores and trend direction across consecutive test images.
- Generated visualizations: panel plots, mapped anomaly scores, and masks to communicate severity and change intuitively.
- All experiments were run on the MVTec AD dataset, adapting static defect images into synthetic temporal sequences.
Mathematical scoring:
[ A_{score} = 0.5 \cdot (1 - \text{SSIM}) + 0.3 \cdot \frac{D_{hist}}{1000} + 0.2 \cdot \frac{D_{HOG}}{50} ]
Dataset Used
- MVTec AD is a collection of over 5,000 high-res industrial images (e.g., screws, nuts, fabrics), provided for research in unsupervised anomaly detection and industrial inspection.
- Each dataset folder includes defect-free images ("train") and a variety of defect types in "test" images. Defects include real cases like scratches, dents, contaminations, and deformations [web:12][web:23].
- While not true video or frame-by-frame time-series, by grouping test images of the same defect type, we simulate progressive defect evolution—demonstrating the pipeline’s time-series potential.
Challenges we ran into
- Aligning normal and test images to minimize false detections (static dataset limitations).
- Maintaining smooth, fast performance in Colab while processing large, high-res industrial images.
- Calibrating anomaly scores/thresholds for different types of manufacturing defects with limited ground truth.
Accomplishments
- Achieved more robust detection by combining multiple image features (ensemble > single-method).
- Successfully tracked both the growth and reduction of defects across a pseudo-timeline of images.
- Created visualizations that clearly communicate temporal change, severity, and spatial position for manufacturing QA.
What we learned
- Ensemble feature engineering detects a wider spectrum of defects and changes than any single method.
- Time-based scoring trends provide actionable insights—can trigger early
Log in or sign up for Devpost to join the conversation.