Client Background

The client is a leading global technology and semiconductor manufacturing company. Operating at the forefront of Industry 4.0, they maintain extensive, highly automated fabrication facilities and advanced packaging lines. Their operations require absolute precision, utilizing complex machinery to produce cutting-edge logic circuits and sensor arrays.

Challenge

Maintaining zero-defect manufacturing required deploying automated visual inspection deep within densely packed production lines and directly onto high-speed robotic arms. This posed severe physical, thermal, and security bottlenecks:

Space Constraints: Traditional cameras rely on refractive glass lenses, which require a minimum focal distance and add significant physical bulk. These cameras simply could not fit into the narrow gaps of the client's manufacturing equipment.

The Miniaturization Paradox: Attempts to use miniature traditional cameras resulted in tiny sensors with poor light collection, compounded by overheating from focus actuators and associated control electronics packed into confined enclosures. In a high-speed industrial setting, these miniature cameras could not reliably capture the high-fidelity images required for automated defect detection.

Security Concerns: Standard cameras capture plaintext, easily interpretable video feeds. In a highly classified semiconductor facility, any network breach of these cameras could expose proprietary assembly techniques and intellectual property.

Yodo Labs' Solution

To address these physical, thermal, and security constraints, the client partnered with Yodo Labs. We delivered a custom, enterprise-grade machine vision architecture: the Lensless Camera.

Based on the academic research and inventions of Yodo Labs' founder, Xiuxi Pan, PhD, this solution entirely discards traditional glass lenses. By replacing multi-element lens assemblies with a single ultra-thin optical mask and a trained AI reconstruction engine, Yodo Labs provided a software-defined camera that is compact, privacy-enhancing, and tailored for harsh industrial environments. [1]

How It Works

The Lensless Camera involved a tight hardware-software co-design process, integrating Yodo Labs' algorithms with the client's manufacturing capabilities:

Monolithic Hardware Integration: To ensure the camera could withstand factory vibrations without losing alignment, Yodo Labs collaborated with the client to monolithically integrate the optics. The custom optical mask and CMOS image sensor were fabricated as a single solid-state module measuring less than a millimeter in thickness. Because the optical path is entirely passive — no focus actuators or associated control electronics — the design eliminates the overheating problems that plagued the client's earlier miniaturized cameras. And because the mask modulates light across the full sensor area rather than concentrating it through a miniature aperture, the system maintains sufficient signal even at this reduced form factor.

Mask-Based Optical Encoding: Instead of focusing light to a single point, the mask modulates incoming light, casting a complex, pseudo-random shadow pattern (Point Spread Function) globally across the sensor. To the human eye, the raw captured data looks like unintelligible static. This visual obfuscation adds a layer of privacy: the encoded data cannot be casually interpreted, though it should not be treated as a cryptographic security guarantee, since reconstruction may be possible if the decoding model or system parameters are compromised.

Vision Transformer (ViT) Decoding: Existing CNN-based reconstruction approaches showed limited effectiveness on lensless data, where task-relevant features are dispersed globally across the sensor rather than localized. Yodo Labs deployed a Vision Transformer algorithm utilizing a Multi-Head Self-Attention mechanism to evaluate the entire encoded pattern simultaneously, enabling high-quality image reconstruction suitable for operator review and inspection workflows when visual verification is required. [1]

Reconstruction-Free Sensing: For tasks where a human-readable image is unnecessary, the system can also perform object recognition and defect classification directly on the raw encoded pattern, bypassing the reconstruction step entirely for lower latency. In the client's deployment, both modes are available: direct classification on encoded data for the real-time automated inspection pipeline, and on-demand reconstruction performed locally on authorized operator workstations for review and audit. Only encoded data traverses the production network by default.

Results

The Lensless Camera delivered concrete improvements to the client's automated quality control:

• Form factor: The sub-millimeter imaging module was deployed in inspection points where no traditional camera could physically fit, expanding automated coverage to previously manual-only stations.
• Thermal stability: Passive optics eliminated the overheating failures that had blocked earlier miniaturized camera trials, enabling continuous operation without thermal throttling.
• Visual privacy: Raw sensor data is not human-interpretable, reducing the exposure risk from network-connected cameras in classified production areas.
• Inspection throughput: Real-time defect detection on the production line, with direct classification for automated inspection and reconstructed-image review available on demand.

Yodo Labs executed this deployment end-to-end: from requirement definition and hardware-software co-design, through PoC and MVP development, to factory integration and ongoing operational support.

References

1. Pan, X. et al., Lensless Imaging Transformer. github.com/BobPXX/Lensless_Imaging_Transformer