Hey everyone!
I'm currently working on a project called VisionQC, an AI-based system to detect defects in industrial products, starting with bottles.
The project is still a work in progress. I'm not here to showcase a finished product or claim that I've built something revolutionary. I actually want to get feedback from people who've worked in computer vision, machine learning, or industrial quality inspection before I take the project any further.
A little about me
I have 3+ years of experience in software development, primarily in full-stack development, along with some experience integrating AI into applications.
I've also worked on big data projects, which gave me experience with Python and technologies like Kafka, Elasticsearch, Kibana, and other tools in that ecosystem.
I'm now pursuing my MSc in Artificial Intelligence and trying to get beyond simply integrating AI APIs into applications by understanding how ML models actually work, training them, evaluating them, and building projects from scratch.
I'm still a newbie when it comes to the deeper ML and computer vision side of things, so I'm learning as I go.
What is VisionQC?
The basic idea is to explore how computer vision can help identify defective products during industrial quality inspection.
I'm using the bottle category from the MVTec Anomaly Detection dataset and experimenting with a convolutional autoencoder using PyTorch.
The current approach is to train the model on normal bottle images, let it reconstruct those images, and use reconstruction errors to identify potential defects.
I'm also exploring how well the model can identify the actual locations of defects, rather than just classifying an entire bottle as defective or normal.
The project isn't complete yet, and I'm still figuring out what works, what doesn't, and what could be improved.
Here's where I'd love your honest feedback
I'd genuinely appreciate it if you could roast the idea or point out things I'm overlooking.
Does using an autoencoder make sense for this problem, or would you approach it differently?
Are there better approaches I should explore for anomaly detection and defect localisation?
What would you focus on to make this more than just a basic academic project?
If you were building this from scratch, what would you do differently?
I'm especially interested in hearing from people who've worked with computer vision, anomaly detection, or real-world ML systems.
I'm sharing this before the project is finished because I'd rather get constructive criticism now than realise later that I could have approached the problem much better.
Here's the GitHub repository: https://github.com/anupam-17/VisionQC
Thanks! All honest feedback is welcome.
Top comments (0)