Team RJAC Edge AI Innovation Challenge 2024

Image Acquisition

• OV7670 camera captures product images, transmitted via B-CAMS-OMV module to STM32H7471-DISCO

Deep learning model segments image [Binary Segmentation]

• Black pixels: Defect areas
• White pixels: Non-defect regions
• Enables precise defect localization and shape analysis

Model evaluation and optimization [Defect Analysis]

• Compare U-Net, FCN, DeepLab architectures
• Optimize for STM32 using STM32Cube.AI

Automated Response

• Real-time actuator control based on segmentation output for defect handling

Deep Learning Architectures

Below are the models which choosen and compared for our use case and each model has a branch with all specifications regarding training and inference

• U-Net
• Mask R-CNN
• DeepLab
• Fully Convolutional Networks (FCN)
• Gated-SCNN
• SegNet
• PSPNet (Pyramid Scene Parsing Network)
• HRNet (High-Resolution Network)

Model Output

Model inference on STM32 hardware

Credits💫

---

GitHub @RionDsilvaCS · Linkedin @Rion Dsilva

GitHub @Aniesh04 · Linkedin @Aniesh Reddy Gundam

GitHub @CharanArikala · Linkedin @Sai Charan Arikala

GitHub @Jahnavi0504 · Linkedin @CH V N S Jahnavi