Video has become the core data foundation for security prevention and control as well as operational decision-making. However, behind the flood of data lies a "storage dilemma": research shows that about 70% of surveillance videos are static or low-value images. Mixing them with effective content not only wastes storage resources but also drives up the comprehensive costs of hardware procurement and server room operation and maintenance.

As a driving force behind technological evolution in the security industry, Hikvision has been anchoring on "unlocking the core value of video" for over two decades, continuously leading industry innovation. The recently released Guanlan encoding technology, leveraging profound industry insights and AI-driven precise encoding logic, breaks the industry's mantra that "image quality and storage efficiency cannot be achieved simultaneously", offering a storage optimization solution that "reduces costs without compromising efficiency" for various industries.

Technical accumulation: Over 20 years of project practice refining core capabilities

Hikvision's exploration of video value stems from years of technological iteration and project practice:

In 2015, it took the lead in integrating deep learning algorithms into front-end devices, ushering in the era of AI in security.

In 2017, the AI Cloud three-tier architecture was introduced, addressing the pain points of "dispersed computing power and insufficient collaboration" and providing strong support for the implementation of AI.

In 2018, we launched an AI open platform to assist customers with no algorithm foundation in developing industry-specific algorithms, thereby promoting the deep integration of video perception technology and industrial applications.

In 2021, Hikvision officially invested in the research and development of the Guanlan large model, establishing a three-tier system of "foundation-industry-task" to lay a solid foundation for video understanding algorithms.

From the basic perception of "seeing clearly", to the precise presentation of "seeing distinctly", and then to the deep interpretation of "seeing comprehensively", Hikvision's core evolution focuses on five dimensions:

Upgrade from image structured labeling to deep semantic understanding of video.

· Iterate from single-scene discriminative models to visual multimodal large models.

· Shift from distributed computing power at the edge to large-scale computing power deployment at the cloud and edge.

· Expand from core target recognition of people and vehicles to cover a full range of long-tail objects.

Upgrade from tabular filtering to semantic human-computer multimodal interaction.

Over the years, Hikvision has served millions of customers globally, deeply exploring various scenarios such as transportation, finance, industrial parks, and healthcare. It is this technology iteration that "originates from practice and is applied to practice" that has laid a solid foundation for the birth of Guanlan encoding technology.

Technological upgrading: AI empowers to break bottlenecks

The mainstream traditional encoding technology in the industry can achieve 30% to 70% storage optimization by dynamically adjusting the compression strength, but it has obvious limitations: the recognition based on the motion state of objects is prone to misjudging non-critical moving objects (such as leaves blown by the wind, flying birds, etc.), and changes in lighting can also affect the effect. It is only suitable for simple static scenes such as warehouses and corridors.

Hikvision's Guanlan encoding technology integrates the deep understanding capabilities of the Guanlan large model with pixel-level segmentation technology, establishing a full-chain optimization system of "intelligent recognition - precise ROI segmentation - differentiated encoding", achieving dual synergy between storage and image quality.

Actual measurement data: storage efficiency has been significantly improved

Powered by large models, digital intelligence is upgraded

With the support of our self-developed Guanlan visual large model, we can accurately analyze high-value key targets such as people, vehicles, and non-motorized vehicles. The recognition accuracy is significantly improved compared to traditional algorithms, and it can support the recognition of up to 64 targets simultaneously.

Through refined ROI protection and segmentation technology, the foreground object and background area are precisely separated: the foreground is encoded conventionally to ensure complete details, while the background is compressed efficiently to reduce storage occupation. The final output is a compliant bitstream that meets standards, balancing image quality and efficiency. While ensuring comparable quality to non-targeted human-machine interactions, the bit rate is saved by 20% to 90%.

Scene perception, dynamically adapting to all scenarios

Hikvision's Guanlan encoding technology centers around "scene perception" to construct a dynamically adaptable intelligent encoding system, relying on the collaborative operation of dynamic and static perception:

Dynamic perception: Capture the motion amplitude and detail density of the scene in real-time, allocate resources based on the percentage of the maximum bit rate MaxBps, and ensure lossless image quality in complex scenes.

Static perception: For still or low-motion images, repeated frame coding is employed, with each single frame requiring only tens of bytes, maximizing storage cost optimization.

Taking the scene of the park entrance and exit as an example: during the morning rush hour, restore details at high bitrates; during the evening, balance image quality and efficiency at 50%; during the early morning, compress to maximize storage savings at 10%. Through dynamic adaptation based on scene complexity, achieve precise allocation of encoding resources.

Guanlan coding saves money, energy, and space

Taking a 2000-channel 1080P@2Mbps system with 90-day storage as an example, compared to traditional encoding, the Guanlan encoding solution achieves a 60% reduction in the number of hard drives, a 60% reduction in server room space, and a 50% reduction in electricity costs over five years. This significantly reduces project costs and truly achieves "cost reduction without efficiency loss".

From the implementation of deep learning to the application of fully structured data, and then to the release of Guanlan encoding technology, Hikvision has been continuously committed to "making video generate greater value" and promoting the evolution of the security industry from "passive recording" to "active intelligence". Through the deep integration of AI and encoding, video data has been transformed from "massive redundancy" to "precise and efficient", injecting new momentum into the intelligent upgrade of various industries.

Currently, Guanlan encoding technology has been applied to Hikvision's front-end and back-end products, and will be released in the near future. Stay tuned!