Inspiration
In industries like manufacturing, quality inspection is often a manual, repetitive, and error-prone process. Tiny defects or surface irregularities can go unnoticed, leading to costly reworks and compromised product quality. The inspiration behind VizInspect came from the need to automate this process and to build a system that can see, compare, and reason over visual differences just as a human would, but with far greater speed, consistency, and precision. We envisioned a framework capable of automatically detecting and classifying visual changes in manufactured products, setting the foundation for a scalable, general-purpose visual inspection engine.
What it does
VizInspect analyzes pairs of product images captured at different times or from different manufacturing batches and identifies regions of visual difference. It not only detects where a change has occurred but also classifies the type of change such as defect, deformation, or environmental variation. The system quantifies differences using metrics like SSIM, MSE, and deep feature similarity, and visualizes them through heatmaps and overlay masks. The result is an explainable and data-driven inspection process that enhances quality control and reduces dependency on manual checks.
How we built it
We built VizInspect as a modular framework combining classical computer vision and deep learning:
Image Preprocessing: Image alignment using ORB feature matching and homography, followed by lighting and color normalization. Change Detection: Pixel-level difference analysis using SSIM, MSE, and absolute difference maps. Deep Feature Extraction: Utilized pretrained Mobile-NetV2 for semantic-level comparison through cosine similarity. Classification: Developed a rule-based severity classifier to categorize differences as No Change, Minor Change, or Significant Change. Visualization: Built an interactive Streamlit dashboard for displaying side-by-side comparisons, difference heatmaps, and severity metrics.
Challenges we ran into
- Balancing accuracy and efficiency to ensure real-time analysis.
- Creating a generalized pipeline adaptable to multiple defect types without overfitting to specific products. ## Accomplishments that we're proud of
- Successfully built a working prototype that can automatically detect and classify changes in manufacturing product images.
- Achieved high accuracy and visual interpretability using a hybrid approach that fuses classical and deep learning techniques.
- Developed a user-friendly dashboard that transforms complex visual analytics into intuitive insights. ## What we learned Through VizInspect, we learned the power of integrating classical image processing with deep feature extraction to balance interpretability and performance. We also gained insights into designing efficient visual inspection pipelines, understanding similarity metrics and transforming raw visual data into actionable insights. Most importantly, we learned how explainability plays a crucial role in building trust in AI-based inspection systems. ## What's next for VizInspect Our immediate next step is to expand VizInspect beyond the manufacturing domain into other industries where visual change detection is equally critical, such as infrastructure monitoring, retail brand compliance, and construction quality audits. We plan to enhance the system with integration of real-time video analysis, and eventually develop a cloud API for enterprise-scale deployment. The long-term vision is to make VizInspect a universal engine for intelligent visual difference detection across domains.
Log in or sign up for Devpost to join the conversation.