How HIK Cameras Process Video: ISP, SoC and Encoding

Two cameras can share a resolution, a lens and even a sensor, yet look and record very differently. The difference is usually the video processing chip: the system-on-chip that turns sensor data into the stream your recorder stores.

This guide explains what that chip does, how the ISP, codec and NPU interact, and which specifications actually matter when specifying cameras for a project.

1. A camera is a small computer

A network camera contains an image sensor, a lens, DRAM, flash and a video SoC. The SoC runs the operating system, the image pipeline, the encoder, the analytics engine and the network stack.

The blocks that matter for image quality and bandwidth are the ISP, the video encoder and, increasingly, the neural processing unit. Memory and storage sit behind them and quietly limit what the camera can sustain.

Because these blocks share one chip, they compete for power and bandwidth. A camera that adds a second stream, a smarter analytics model or a higher frame rate is not adding a feature in isolation; it is taking resources from somewhere else.

The data path runs from the sensor over a serial link into the SoC, through the ISP, into the encoder, and out over Ethernet. Analytics taps the same frames, which is why detection and encoding settings interact.

Two cameras with the same sensor can therefore differ in sharpness, noise, bitrate and detection reliability. The sensor sets the ceiling; the processing chain decides how much of that ceiling the viewer actually gets.

2. The ISP: where the image is made

The image signal processor takes raw Bayer data and produces a viewable frame. It applies black-level correction, lens shading correction, demosaicing, white balance, gamma, sharpening and noise reduction.

Wide dynamic range is the ISP feature buyers notice most. A camera facing a bright entrance needs tone mapping that keeps both the doorway and the interior readable; poor WDR shows as a white doorway or a black interior.

Temporal noise reduction uses information across frames. It cleans up low-light video, but aggressive settings smear moving objects, which matters when the purpose is identification rather than a pleasant picture.

The ISP also handles day/night switching and IR-cut control. A slow or unstable switch produces colour frames at dusk that are noisy and unusable, a common complaint that has nothing to do with the sensor.

Colour science is part of the same block. Saturation, contrast and white balance are vendor choices, and two cameras can render the same red jacket differently. For evidence work, neutral colour and accurate skin tones matter more than vivid marketing images.

Lens shading correction is easy to overlook. Cheaper lenses fall off at the corners; the ISP compensates, but heavy correction amplifies corner noise at night.

3. The encoder: bandwidth and storage

The encoder compresses frames into H.264 or H.265. H.265 typically reduces bitrate by 40–60% at the same perceived quality, which is why it dominates new projects.

Smart codecs go further by raising the I-frame interval and reducing detail in static areas. On a quiet scene the saving is large; on a busy street the encoder spends bits to keep moving objects sharp, and the saving shrinks.

Three settings drive the real bitrate: resolution, frame rate and scene complexity. Dropping from 25 to 15 frames per second can cut bitrate by roughly a third with little loss for corridors and perimeter views.

Sub-streams are encoded separately at lower resolution. They are what mobile apps and multi-view walls use, and they are the difference between a usable remote view and a spinning loader.

Rate control matters as much as the codec. Constant bitrate keeps the stream predictable for networks and storage; variable bitrate spends bits where the scene needs them. Recorders usually prefer the predictable option, while image quality prefers the adaptive one.

The codec is defined in standards such as ITU-T H.265, but the implementation quality varies by chip. A well-tuned H.264 encoder can beat a poor H.265 one at the same bitrate.

4. The NPU: analytics on the edge

Human and vehicle classification runs on a neural processing unit inside the SoC. It filters out rain, trees and headlights that would trigger pixel-based motion detection.

NPU capability is quoted in operations per second, but the useful question is what the camera can run at once: how many detection types, at what resolution, on how many streams, at what frame rate.

Analytics also depends on pixels on target. A person forty pixels tall at the far end of a car park will not be classified reliably, whatever the NPU specification says.

Edge analytics reduce recorder load and network traffic, but they are not free: they consume SoC headroom that would otherwise go to encoding quality or a second stream.

A typical analytics pipeline runs detection, classification and tracking. Detection finds moving regions; classification labels them; tracking keeps an identity across frames so one person does not generate twenty events.

Interoperability matters when cameras and recorders come from different vendors. ONVIF profiles define how analytics metadata is exchanged, and a recorder that ignores those events will show the camera as a plain video source.

Privacy masking runs on the same path. Masks applied after encoding are less reliable than masks applied in the ISP before encoding, because the underlying pixels may still exist in the raw stream.

5. Memory and storage

DRAM bandwidth limits how many streams and analytics the SoC can sustain. A camera with plenty of compute but thin memory will drop frames when analytics and a second stream run together.

On-board storage uses eMMC or NAND flash, or a microSD card. Cards are convenient but have limited write endurance; continuous recording on a consumer card can wear it out within a year or two.

Where a camera records to a card as a backup, choose a high-endurance card and monitor its health. Where it records only events, a standard card is usually adequate.

A simple endurance estimate helps. A 64 GB card rated for 3,000 write cycles can absorb roughly 190 TB of writes before wear-out. A single 4 MP camera at 4 Mbps writes about 1.3 TB per month, so continuous recording fills that budget in a few years, not decades.

Network storage shifts the problem to the recorder. The camera still buffers frames in DRAM, so a weak memory subsystem shows up as dropped frames during playback, not as an obviously broken live view.

6. Power, heat and PoE

Encoding and analytics consume power, and power becomes heat in a sealed housing. A camera that runs cool at 1080p may throttle or reboot at 4K with analytics enabled in direct sun.

PoE budgets are often calculated from the camera's idle draw. Add the IR LEDs, the heater where fitted, and the extra load when analytics runs, then leave headroom on the switch.

Heat also affects image quality: sensor noise rises with temperature, so a well-designed housing and a sensible bitrate often matter more than another megapixel.

These are planning figures. Measure the actual draw at night with analytics enabled before finalising a switch budget.

7. What to check when specifying cameras

For component-level sourcing — image sensors, video SoCs, PoE controllers, memory and storage — Chip-seek lists available electronic components with stock and pricing, and its component catalogue is a useful reference when a project needs a specific part rather than a finished camera.

8. FAQ

Does a bigger NPU mean better analytics? Not automatically. Resolution on target, model quality and tuning matter as much as raw compute.

Is H.265 always better than H.264? At the same quality it usually saves bandwidth, but compatibility and recorder support still matter.

Why does bitrate rise at night? Noise is expensive to encode. A noisy image needs more bits for the same perceived quality.

Can I add analytics with firmware? Sometimes, if the SoC has an NPU and the vendor releases the feature, but do not assume it.

Do smart codecs affect evidence quality? They can. Test the actual scene before relying on aggressive settings for identification.

How many streams can a camera run? Typically two or three, depending on resolution, codec and SoC headroom.

Is microSD reliable for 24/7 recording? Only with high-endurance cards, and even then health should be monitored.

Does heat affect image quality? Yes. Sensor noise increases with temperature, especially in compact housings.

Why do two 4 MP cameras need different storage? Bitrate differs with codec tuning, frame rate, scene complexity and noise at night.

Can I compare cameras by datasheet alone? Only partly. Test the actual scene at night; processing differences rarely show on a spec sheet.

Recommended HIK cameras

Cameras with current ISP, codec and analytics platforms. Confirm the analytics types and stream limits for your project.