Introducing a custom face recognition software solution

Innowise has developed an advanced face recognition solution to enhance security and streamline identification processes.

Customer

Industry

Retail industry

Region

Client since

2022

Our client is an Australian software development and IT consulting company specializing in delivering tailored IT solutions for the retail industry. With a strong track record, they have successfully conceptualized, designed, and launched digital solutions across various retail categories, including general merchandise, apparel, and grocery.

Operating on a global scale — ranging from multinational retail corporations to independent store owners — this company offers a suite of highly scalable products and services.

Challenge

Enhancing facial recognition accuracy for the retail industry

At first, the task seemed straightforward: develop a face recognition solution that could reliably identify individuals in real time for retail environments. But anyone familiar with real-world video feeds knows they’re rarely perfect.

The main challenge was inconsistent video quality. Retail settings are unpredictable — cameras often capture footage in poor lighting, from awkward angles, and amidst constant movement. As a result, faces appeared blurry, shadowed, or distorted, making it difficult for the system to detect and align critical facial features like the eyes, nose, and mouth.

In some cases, uneven lighting obscured facial details, while in others, the combination of blur and shadows rendered traditional recognition methods ineffective. These weren’t occasional issues — they were the everyday conditions our team had to address.

To overcome this, we needed more than advanced algorithms. Our engineers had to design a system capable of processing imperfect, messy inputs — extracting meaningful data from low-quality, inconsistent video to deliver clear and actionable results. Simply put, the custom facial recognition software solution had to work with real-world challenges, not fight against them.

Solution

Enhancing retail security and efficiency with a face recognition system

Our team’s approach to custom facial recognition software development services focused on three critical goals: accuracy, real-time processing, and adaptability to low-quality inputs.

Reliable face detection and recognition

At the core of the solution, we integrated advanced deep learning algorithms to ensure precise face detection and recognition, even in demanding conditions such as poor lighting, unusual angles, and low-resolution inputs.

Face detection and alignment

We used RetinaFace for its speed and accuracy in face detection, particularly excelling in low-resolution images and challenging lighting conditions. We opted for MediaPipe’s facial landmark detection to identify and align critical features such as the eyes, nose, and mouth. As a result, the system could handle diverse inputs with greater stability and accuracy. This enabled consistent preprocessing of faces, even under misalignment or unusual angles.

Accurate face recognition models

For face recognition, we utilized the ArcFace technique, known for its strong performance in generating discriminative facial embeddings. To optimize accuracy for retail environments, the team fine-tuned the backbone model using domain-specific data with targeted augmentations, including simulated blur and angle distortion. As a result, the system achieved 85–90% face recognition accuracy in challenging conditions and maintained over 95% accuracy with high-quality inputs.

Enhancements for image accuracy and system performance

Surveillance footage often comes with imperfections, so, as part of our custom face recognition development services, we developed a robust image preprocessing pipeline to clean up the inputs before recognition.

One of the key breakthroughs we brought in was eye localization. The integration of MediaPipe pipelines improved the system’s ability to detect eye pupil centers. This greatly improved face alignment and stabilization, enabling us to filter out the frames that were either poorly captured or misaligned. In this way, only clean and high-quality frames went to recognition, hence improving the overall system accuracy.

Efficient batch image processing

We needed to deal with huge amounts of video data, so we created a batch image processing module using PyTorch and MediaPipe.

We also developed a module to extract and process images from video feeds in bulk, thereby saving time and reducing manual effort. The optimized system handles high volumes of data seamlessly, even in busy retail environments.

Integration with a CCTV system

The integration of custom face recognition software with a closed-circuit television (CCTV) system involves combining advanced facial recognition technology with the existing surveillance infrastructure. This integration reconfigures surveillance capabilities, allowing for precise real-time identification of individuals within stores or warehouses. Such a system tightens security measures against unauthorized access and optimizes employee management through attendance tracking and monitoring of work behavior. With such a holistic approach, the environment of any retail outlet becomes much safer and efficient for operation.

Now, the custom face recognition solution accesses live video feeds from CCTV cameras and employs PyTorch and MediaPipe-based algorithms to detect and analyze unique facial features, such as the shape of the eyes, nose, and mouth.

Using person re-identification (Re-ID) models, the system tracks individuals from one camera to another, even with occlusion or while moving from one zone to another. Combined with streamed frame-level processing powered by PyTorch’s inference capabilities, the system also supports real-time recognition with sub-200ms latency, even across multiple live streams.

Technologies

Back-end

.NET 6, C#, Python

Cloud

AWS (Kinesis Video Streaming, EC2, EKS, ECR, S3, Glue)

DevOps

GitHub Actions, Kubernetes, Nginx

Machine learning

PyTorch, MediaPipe

VCS

Git, GitHub

Process

We began with intensive workshops to understand the project goals and the challenges at hand — like
handling poor video quality, enabling real-time processing, and ensuring the system could scale. Our team performed a
detailed audit of the client’s CCTV setup, assessing camera types, frame rates, and video quality to make sure the
custom facial recognition software would work reliably in real-world conditions.

System design and architecture planning

Next, our team designed a scalable, distributed architecture capable of processing multiple live
streams simultaneously. Each part of the system — face detection, preprocessing, and recognition — was built as an
independent component, ensuring smooth data flow and fault tolerance. We also mapped out integration points to connect
the solution with the client’s existing CCTV network.

Development and implementation

We followed an Agile development approach, delivering results in stages and gathering regular feedback
to refine the system. Here’s how we tackled each critical area:

Face detection and alignment: Our engineers built and optimized the detection pipeline to
stabilize
incoming video, addressing challenges like inconsistent lighting, unusual angles, and movement.
Image preprocessing: Our team enhanced video frames with filtering and stabilization,
ensuring only clean,
usable frames moved forward for recognition.
Recognition models: Using real-world video samples, we fine-tuned the recognition pipeline
to accurately
identify faces, even under extreme angles or partial obstructions.
Batch processing optimization: To manage large video volumes efficiently, our optimized
batch processing
workflows for archived footage while reducing the load on real-time operations.

At every sprint, we conducted rigorous testing and performance monitoring to address bottlenecks and
support consistent progress.

Testing and validation

Our QA specialists put the system through its paces to validate its performance under real-world conditions:

Functional testing: Each module was tested individually and as part of the full pipeline.
Stress testing: Our team evaluated the system’s ability to process multiple camera streams under heavy loads.
Real-world validation: Using actual retail video feeds, we verified that the system delivered consistent results even with poor lighting, motion blur, or low resolution.
Edge case testing: Our team tested challenging scenarios like partial faces and extreme angles to guarantee robustness and reliability.

Throughout testing, we tracked performance metrics — accuracy, speed, and frame rejection rates — and fine-tuned the system for optimal results.

Deployment and integration

Once the custom face analysis software was ready, our team deployed it into the client’s production environment with minimal disruption. The system was configured to process live video streams and integrate effortlessly with the existing CCTV infrastructure. To ensure a smooth rollout, we also provided training sessions and detailed documentation for the client’s team.

Post-deployment maintenance and support

As part of our role as a custom facial recognition software development company, we provide continuous updates and support to improve system efficiency and scalability.

Team

Business Analyst

Project Manager

ML Engineer

Back-End Developer

Front-End Developer

Results

Enhanced security and streamlined face identification

Our team delivered a facial recognition system that successfully addressed key real-world challenges, including low resolution, poor lighting, and motion blur. Through careful design and optimization, we improved operational efficiency by 70%, reducing employee verification time from 20 seconds to under 5 seconds per person.

Our engineers made sure the system could handle demanding workloads by implementing efficient processing pipelines. As a result, it now processes thousands of faces per second across multiple video streams. Utilizing GPU-optimized AWS infrastructure and fine-tuning performance, we were able to keep things running smooth and consistent, even during peak retail hours.

Our efforts also strengthened security outcomes. The real-time alert mechanisms developed by our team enable the system to generate instant notifications for unauthorized individuals. As a result, security response times decreased by 40%, allowing on-site teams to act more quickly and improving overall situational awareness.

Reliability was a key focus throughout the project. Our team’s optimizations ensured 99.9% uptime and provided uninterrupted operation for critical processes like access control and live monitoring. Seamless integration with the client’s existing systems further contributed to a 20–25% reduction in security-related incidents, helping retailers create safer and better-managed environments.

Overall, the solution proved to be fast, accurate, and scalable. It not only optimized security but also simplified attendance management and improved day-to-day operational workflows, delivering tangible results for retail environments.

Project duration

May 2022 - October 2022

99.8%

verifications under 200 ms per frame on GPU-optimized AWS instances

85–90%

face recognition accuracy in dim lighting or motion blur

50%

fewer manual interventions with reliable automated identification

Related cases

Contact us!

Book a call or fill out the form below and we’ll get back to you once we’ve processed your request.

What happens next?

Having received and processed your request, we will get back to you shortly to detail your project needs and sign an NDA to ensure the confidentiality of information.

After examining requirements, our analysts and developers devise a project proposal with the scope of works, team size, time, and cost estimates.

We arrange a meeting with you to discuss the offer and come to an agreement.

We sign a contract and start working on your project as quickly as possible.