JOB SUMMARY
We are seeking candidate in the field of vision AI to enhance the performance of our existing vision-based defect detection system in manufacturing. The candidate will focus on improving the speed and accuracy of the system while minimizing false positives and false negatives during inference. This role involves analyzing the current end-to-end process, identifying bottlenecks, and implementing optimizations to achieve higher efficiency and reliability in defect detection.
JOB RESPONSIBILITIES
Process Analysis and Optimization:
· Study the existing vision AI pipeline, from image capture to inference on embedded hardware.
· Identify inefficiencies or areas for improvement in model architecture, preprocessing, or hardware integration.
Model Performance Enhancement:
· Reduce false positives and false negatives through techniques such as threshold tuning, data augmentation, and advanced training methods.
Hardware Deployment:
· Refine model deployment on embedded devices by leveraging techniques like quantization, pruning, and knowledge distillation to improve performance within resource constraints.
Data Management:
· Analyze and improve training datasets by including diverse and high-quality samples to enhance model generalization.
· Implement feedback loops to continuously update datasets with edge cases and anomalies encountered during production.
EDUCATION and EXPERIENCE
Master’s degree in Computer Science, Electrical Engineering, or a related field. A Ph.D. is a plus.
JOB-RELATED SKILLS & PRE-REQUISITES
· Proficiency in Python and deep learning frameworks such as TensorFlow or PyTorch.
· Expertise in computer vision techniques, particularly CNNs, for defect detection tasks a plus
· Experience with model optimization techniques like quantization, pruning, mixed precision training, and
knowledge distillation a plus
· Strong understanding of metrics like precision, recall, F1-score, and their trade-offs in balancing false positives and negatives.
· Knowledge of advanced strategies such as ensemble learning or dynamic thresholding to improve detection accuracy a plus
