All cases

Edge AI image reconstruction solution for metalens-based camera systems

Client
A deep tech startup developing metalens-based camera systems for consumer electronics
Industry
Consumer electronics, computer vision, imaging hardware
Services
AI consulting, Edge AI development, PoC development
Tech stack
Qualcomm AI Stack, PyTorch, ONNX Runtime, NVIDIA TensorRT, Python

Challenge

A deep tech startup working on the next generation of smartphone and IoT cameras wanted to assess the feasibility of replacing bulky traditional lenses with ultra-thin metalenses. Their system employs a film of microscale metalenses attached to a complementary metal-oxide-semiconductor (CMOS) sensor to generate multiple sub-images (M x N grid) and reconstructs them into a single high-quality image using custom AI algorithms. To accelerate time to market, the client tapped into ITRex’s electronics prototyping and edge AI development services. We became their primary technology partner, providing guidance on platform selection, optimizing edge AI performance, and creating a compelling proof of concept (PoC) and investor pitch. Here’s how the project unfolded. Limitation of Existing Camera Lenses Traditional smartphone cameras focus light onto a sensor using thick glass or plastic lenses. As imaging demands increase, so does the number of necessary lens elements—hence the bulky, protruding modules that limit design flexibility for consumer electronics brands. Our client set out to change this by creating a next-gen metalens-based system: a wafer-thin film composed of microscopic lenses that sits directly on the image sensor. Rather than forming a full image in one go, this film generates multiple sub-images, with each sub-image capturing the scene from a slightly different angle or optical configuration. A custom AI algorithm then reconstructs a single, high-resolution image from this M x N array.

The advantages are clear:
Ultra-thin form factor—cameras can be as thin as the sensor itself, eliminating the “camera bump” on phones, drones, and wearables
Optical flexibility—different parts of the metalens array can be engineered to capture various wavelengths or focal properties, enabling new types of filters, depth capture, and image enhancements
AI-powered enhancement—proprietary edge AI algorithms speed up the reconstruction of high-fidelity images in real time, even with limited compute resources
However, to build a viable PoC for the edge AI solution and convince investors, the client needed:
An efficient implementation of their neural image restoration pipeline on edge hardware
Advice on selecting the best frameworks and AI acceleration platforms
A compelling technical vision and investor pitch to support fundraising efforts

Solution

ITRex used its expertise in edge AI, computer vision, and embedded systems to help the startup move from idea to investment-ready proof of concept.
Edge AI strategy & platform selection We devised a hardware-agnostic deployment strategy centered on ONNX Runtime. This option provided maximum flexibility, allowing the AI model to run on a variety of platforms while leveraging hardware-specific execution back ends. Due to its popularity among consumer electronics companies, the ITRex team suggested focusing initial validation on the Qualcomm AI Stack. We also validated the solution’s performance on NVIDIA TensorRT, paving the way for scalability and avoiding potential vendor lock-in.
Model adaptation & optimization Our engineers adapted the client’s image reconstruction models (originally built in PyTorch) for edge deployment. We used open-source tooling to convert models to ONNX format and then fine-tuned them for speed and memory efficiency on Qualcomm-based hardware. The team also implemented runtime profiling to compare throughput and latency across platforms.
PoC implementation We provided a small, self-contained proof of concept capable of restoring full-resolution images in real time on resource-constrained devices. The pipeline reconstructs the entire scene from multiple metalens sub-images using edge-optimized AI algorithms.
Technical video & investor pitch To support the client's investment outreach, ITRex created a short technical demo video demonstrating image reconstruction in action, as well as an investor-facing deck that clearly explained the technology stack, benefits, and business potential.
Edge AI Image Reconstruction for Metalens Startup
Edge AI Image Reconstruction for Metalens Cameras

Impact

Through collaboration with ITRex, the startup was able to:
Validate the core technical concept with a fully working PoC on edge hardware
Reduce time to market by choosing well-documented open-source deployment platforms
Clarify scaling paths and hardware compatibility for future production
Impress early-stage investors with a strong technical pitch and visual demo
Maintain flexibility by creating a system that can switch from open-source to licensed frameworks as needed
The client is now moving forward with MVP development, using the same optimized stack to improve image quality and add more metalens configurations.

Technicalities

Neural image restoration pipeline The customer initially developed the image reconstruction algorithm in PyTorch. We adapted it for ONNX Runtime to support deployment across edge platforms. The system receives an M x N grid of sub-images from the metalens array and produces a single high-resolution image. Denoising, alignment, and fusion stages are all part of the image reconstruction process, with AI models trained to deal with real-world artifacts.
Edge optimization Our team used quantization and hardware-aware tuning techniques to optimize the ONNX model. To validate its performance, we benchmarked the pipeline across multiple platforms using ONNX Runtime with hardware-specific execution providers. This included using Qualcomm's Neural Network (QNN) Execution Provider on Qualcomm chipsets and NVIDIA's TensorRT Execution Provider on Jetson devices. This method allowed us to directly compare the model's real-time throughput and memory footprint across various edge ecosystems.
ArchitectureModeling framework: PyTorch ● Model export format: ONNX ● Inference engine: ONNX Runtime ● Hardware acceleration back ends: Qualcomm AI Stack (via QNN), NVIDIA TensorRT ● Deployment environment: Qualcomm Linux ● Tooling: Python-based wrappers, real-time monitoring scripts ● PoC deliverables: Integrated camera + metalens module, test image dataset, inference engine, demo video
Edge AI Image Reconstruction

Latest projects