As surveillance image recognition technology matures, it not only strengthens internal security and protects vital physical assets, but also enables real-time incident detection and monitoring. In addition to theft and vandalism detection, AI can be used to analyze vast amounts of surveillance footage to produce useful insights such as crowd flow, vehicle identification, and occupancy tracking.
Synology has dedicated years to the development of comprehensive image recognition products and services. From the DVA series, which provides backend AI analysis for third-party video streams, to Synology Cameras equipped with edge computing abilities, Synology has outfitted enterprise-level businesses worldwide with surveillance solutions tailored to their unique needs. Synology’s unique deep learning algorithms have achieved over 97% accuracy in the National Institute of Standards and Technology (NIST) facial recognition vendor tests in the USA.
These successes raise several important questions for users of image recognition technology. How are these systems trained? What level of training does an image recognition solution need in order to be considered accurate? Can this accuracy be improved? Let’s explore how Synology makes surveillance systems smarter.
Building Accurate Image Recognition Models
Almost all image recognition technology is based on preexisting underlying models, which are then adapted to specific applications. The accuracy of these systems is therefore heavily dependent on the investment of resources and effort by the supplier during development. Synology’s surveillance team has enhanced the accuracy of their image deep learning models through two key phases: extensive training and application integration.
During the training phase, the team not only gathers publicly-available datasets from the internet, but also collects real-world application data, enhancing model performance. For example, outdoor cameras are trained by capturing the same scene under different conditions (Day, dusk, night, sunny, cloudy, and rainy) to help the model improve recognition accuracy in challenging conditions like overcast or rainy weather.
Synology’s machine learning team continuously studies the latest academic papers, weekly sharing and discussing top-tier and cutting-edge research, applying these findings to model training. The famous YOLO model for object detection, which releases new versions periodically, has taught the team advanced model construction and training techniques.
“Given the vast variability in real surveillance scenarios, no image recognition technology is without limitations; no vendor can achieve 100% accuracy,” mentions Steven Liang, manager of Synology’s machine learning group. The team conducts field tests in numerous scenarios to develop more sophisticated algorithms to improve models.
The biggest challenge lies in unpredictable situations that could lead to AI misjudgments, such as cameras being covered by spider webs or nighttime misidentification of objects that have a human-like silhouette. To address misidentification, Synology feeds more data to the model for finer distinction or adjusts the model’s structure to enhance recognition accuracy.
Customer-Centric Solutions Through Deep Learning
Liang emphasized that enhancing user experience goes beyond just refining deep learning models’ accuracy – it is essential to provide AI detection features that are tailored for diverse scenarios, meeting the needs encountered across a wide variety of industries. “Synology focuses on solutions that meet the real-world operational needs of our customers, helping to reduce their management workload and boost security,” he stated. Synology’s image recognition targets key needs in surveillance systems, such as identifying people and vehicles, and can accurately detect face coverings or track movements.
Traditional motion detection, often triggered by minor changes like wind-blown leaves, can lead to false alarms. In contrast, Synology’s AI-driven person and vehicle detection feature is built on deep learning models. This technology offers higher accuracy, precisely identifying people or moving vehicles in footage and tracking their presence in specific areas. Liang further explained that this deep learning model proactively learns to distinguish between humans and vehicles from other objects in the video, overcoming the limitations of previous motions detection models that relied solely on scene changes, which could be easily triggered by irrelevant objects.
Advanced image recognition enhances security by continuously monitoring for predefined threats, such as unauthorized access by banned individuals or vehicles, and ensuring only staff enter restricted areas. Synology’s system also reduces security workload with alert services, like notifying staff when areas become overcrowded, adding in swift response planning.
Developing Synology Cameras with Edge Computing Technology
Synology has emphasized the importance of real-time analysis as a crucial evaluation metric for effective image recognition architecture, highlighting the significance of edge computing. The advantage of edge computing lies in processing images at their source, achieving minimal delay and saving power without overtaxing backend servers with decoding and computation tasks.
Last year, Synology introduced its Synology Camera, a key product showcasing real-time image analysis capabilities. These cameras are equipped with a Neural Processing Unit (NPU) designed for computing and AI tasks, developed by Synology’s machine learning team. The lightweight AI model enables these compact devices to recognize images, including detecting people, vehicles, and intrusion events. When integrated with Surveillance Station, it allows for more immediate image search, helping security personnel easily identify potential threats, quickly clarify suspicious areas, an follow up with relevant surveillance footage retrieval.
Balancing Technological Advancement with Reliable Management Tools
Besides outstanding video analysis technology, Synology provides a comprehensive one-stop surveillance architecture, focusing on deployment and management features that help businesses easily enhance security.
Surveillance Station is compatible with over 150 brands and 8,400 camera models, allowing businesses the flexibility to continue using older devices or upgrade to new devices according to their needs and budget. For multi-site, multi-server surveillance system management, Synology’s Centralized Management System (CMS) offers central control and batch system updates, enabling IT teams to manage up to 10,000 cameras and 1,000 Synology NAS recording servers.
After ensuring the viability of large-scale management, Synology also considers the long-term ownership costs for businesses. Surveillance Station, a commercial video surveillance software running on Synology NAS, uses a perpetual licensing model. This means businesses only need to make a one-time purchase of hardware and camera licenses to enjoy comprehensive surveillance management features, easily building a long-term sustainable surveillance architecture. As organizations expand, licenses can be purchased as needed for added cameras, or can be transferred from old to new cameras when upgrading, significantly enhancing IT budget flexibility.
With these advantages, Synology comprehensively meets organizational security needs, not only pursuing precision in image recognition technology but also focusing on usability and cost, making it the best choice for enterprises investing in smart architecture.
