The Application of Narrowband Filters in Facial Recognition

Facial recognition technology identifies people by their facial features and is a type of biometric recognition. It involves capturing images or video streams containing faces using an image acquisition device, automatically detecting and tracking the face in the image, and then locating and extracting facial features. By comparing these features, it identifies different individuals. The computational process for facial recognition is massive, and both the initial image quality and the algorithm's performance have a decisive impact on recognition efficiency. Here, we focus on analyzing the narrowband filters used in the image acquisition devices of facial recognition systems. The goal is to help users better understand the role and usage of narrowband filters to select the right technical specifications properly.

Light Source

In the image acquisition devices for facial recognition, the light source commonly uses high-power infrared diodes with wavelengths mainly at 850nm and 940nm. To improve recognition efficiency and the utilization rate of light, consideration for the overall design should begin with the selection of the light source. Although the nominal values of LEDs available in the market are 850nm or 940nm, there are often deviations when measuring the central wavelength of specific LED products. For example, for LEDs labeled as 850nm, the actual central wavelength may be found at 835nm or at 865nm. As the light source in facial recognition systems consists of multiple high-power LED arrays, if each LED’s central wavelength varies, the combined spectral bandwidth will broaden after superposition. A single 850nm LED has a bandwidth of about 50nm, but if the central wavelengths differ, the combined spectral bandwidth of multiple LEDs will become much wider. This is unfavorable for the selection of narrowband filter bandwidth, energy utilization rate, and signal-to-noise ratio enhancement. Therefore, it is required that the central wavelengths of the selected LED light sources be consistent. Additionally, as the working temperature of the LED light source increases, its central wavelength drifts towards longer wavelengths, shifting about 1nm for every 10°C increase. Moreover, as the working temperature rises, the luminous efficiency of the LEDs rapidly decreases. When the temperature reaches about 85°C, the output efficiency of the LED drops to about 50%. Therefore, it is essential that the LED light source has good heat dissipation. Also, when choosing the divergence angle of the LED emitter, a smaller angle is preferable to improve the energy utilization rate of the light source.

Receiver 

In facial recognition systems, CCD image sensors are essentially used as receivers. CCDs are favored for their small size, light weight, low distortion, low power consumption, ability to be driven at low voltage, resistance to impact, vibration, and strong electromagnetic interference. Therefore, they are widely applied in various image acquisition systems. The CCDs used in facial recognition systems are typically silicon-based, with a spectral response range from 400nm to 1100nm, which is also the spectral range that needs to be considered for narrowband filters.

Selection of Bandpass Filter

Narrowband filters are primarily used to isolate interfering light, allow signal light to pass through, enhance useful information, and reduce interference, laying the foundation for subsequent image processing and recognition. Currently, facial recognition is mainly used in various attendance and access control systems. Some are installed in places with dim indoor lighting, and others in brighter areas. Under different circumstances, the intensity of the interfering light varies, and so do the requirements for narrowband filters.

It has been found that people often use infrared glass that blocks visible light and allows infrared light to pass as a filter for isolating interference light, which can certainly be effective to some extent. However, ordinary infrared glass only isolates light from the visible and ultraviolet parts and does not block infrared light. In actual interference light, the spectrum exists from visible to infrared, because the spectrum of sunlight is very broad, and diffused or scattered sunlight is the main source of interference. Therefore, to achieve good anti-interference effects, it is necessary to use narrowband filters. The comparison of the transmission performance between absorptive infrared glass and narrowband filters is shown in Figure below. As can be seen from the figure, no matter what type of infrared glass is used, it only isolates visible light and has no blocking effect on infrared light, while the narrowband filter effectively isolates all interfering light outside the range of the signal spectrum.

Determination of FWHM

The bandwidth of a narrowband filter should neither be too narrow nor too wide; it should be determined in conjunction with the environment and the light source used. An 850nm infrared LED has a bandwidth of about 50nm. When selecting a narrowband filter, the utilization rate of light energy must be considered, so the bandwidth of the filter should not be set too narrow. For LED light sources, a bandwidth of less than 15nm is not quite appropriate. On the one hand, a very narrow bandwidth would reject a significant portion of the strong signal light from the LED. On the other hand, a very narrow bandwidth would make the effective use angle of the filter very small, possibly leading to images that are bright in the center and dark at the edges. Through practical testing, it has been found that when setting the threshold of LED luminous intensity utilization at around 70%, the captured images still have fairly good contrast. Therefore, the bandwidth of the narrowband filter can be chosen to be around 30nm, or 20nm for higher interference resistance requirements. 

Determination of the Center Wavelength of Narrowband Filters

Theoretically, the best choice for the center wavelength of a narrowband filter is to match it with the center wavelength of the selected LED. However, as mentioned earlier, there are two factors that can cause slight adjustments to the selection of the center wavelength of the narrowband filter: the incident angle effect and the heat generation of the LED itself.

Incident Angle Effect

In actual camera operation, the light reflected from the human face reaches the filter in a certain range of angles, such as within ±10°. Therefore, the light incident on the filter is not only 0°, but also between 0° and 10°. When a narrowband filter encounters light incident at an angle, the center wavelength of the narrowband filter will shift towards the shorter wavelength direction. For example, for a narrowband filter with a center wavelength of 850nm when incident at 0°, the center wavelength will shift to 847nm when the incident angle is 10°.

Heat Effect

When the temperature of the LED rises by 10℃, the center wavelength of the LED will shift towards the longer wavelength direction by 1nm. These two influencing factors prompt us to consider the factors of change in the process of use when determining the center wavelength of the narrowband filter. Therefore, the center wavelength of the narrowband filter should be set about 5nm higher than the center wavelength of the LED in advance. This takes into account both the incident conditions at angles from 0° to 10°, and also the case where the center wavelength of the LED shifts upwards due to temperature rise.

Determination of the Cutoff Range

The cutoff range of a narrowband filter is mainly determined by the response range of the receiver itself and the wavelength range of the interference sources in the environment where the receiver is located. The response range of the receiver CCD is 400~1100nm. In face recognition applications, the main interference sources are diffuse or scattered sunlight and surrounding artificial light sources, which span a wide wavelength range from ultraviolet to near infrared. For these two reasons, the cutoff range of narrowband filters used for face recognition can be determined as 400~1100nm.

Determination of the Cutoff Depth

Theoretically, the lower the transmittance within the cutoff range, the better. However, considering the manufacturing cost and actual needs, the cutoff depth should be selected at a reasonable value. In face recognition systems, when the cutoff transmittance of a narrowband filter is less than 1%, the isolation effect of interference light can be significantly reflected. The narrowband filters for face recognition should have a cutoff transmittance of less than 0.5%, and the effect of use is very good. For applications where the intensity of interference light in the environment is particularly strong, we can provide products with higher cutoff depth to meet customer needs.

Determination of Peak Transmittance

In general, everyone thinks that the higher the peak transmittance of a narrowband filter, the better. In most cases, this is correct. However, in face recognition applications, this is not always the case. When a face recognition device is in direct sunlight, the intensity of interference light is very strong, and the interference light with the same wavelength as the signal light is also very strong. This interference light cannot be removed by the narrowband filter. At this time, to improve the anti-interference ability, it is necessary to further increase the incident intensity of the LED light to make the intensity of the signal light several times stronger than the intensity of the interference light. Increasing the intensity of the LED light source is relatively simple to implement, just increase the number of LEDs. However, when the energy of the LED light reaches a certain value, coupled with the energy of the interference light with the same wavelength as the LED, the response of the CCD receiver is easily saturated, causing serious image distortion. Even if the exposure is reduced by software, this problem may not be solved. At this time, the narrowband filter needs to play a certain attenuation role in the signal light band while filtering out the interference light in the cutoff region. According to the actual situation, the peak transmittance of the narrowband filter may be required to be 40%, or 60%, or other values.

Selection of Filter Thickness

Considering the cost factor, the CCD receivers and corresponding lens groups used in face recognition systems are currently mostly ready-made and general-purpose, and are widely used in webcams or mobile phone cameras. The zoom range of these general-purpose cameras is relatively small. If a filter is placed in front of the CCD, it will introduce excessive path difference, causing blurred imaging. When the path difference is small, the image can still be made clear by fine-tuning the focus, but when the thickness of the filter is large, the introduced path difference is also large, and it may not be possible to adjust the focus to compensate, resulting in blurred images. Therefore, many people place the filter in front of the lens of the CCD camera instead of in front of the CCD sensor, because this placement is equivalent to the filter not interfering with the imaging path. However, the filters placed in front of the lens are large in size, inconvenient to install, and not aesthetically pleasing, so many people hope to reduce the size of the filter and place it in front of the CCD sensor, built into the camera, which can save the cost of the filter and does not affect the appearance. To place the filter inside the camera, the filter needs to be very thin. In practice, it is found that filters with a thickness of 0.55mm or 0.7mm are suitable.

Precautions for Wide-Angle Field Shooting

If a wide-angle field of view is required, the narrowband filter needs to be placed as close to the CCD or CMOS as possible, which means it needs to be built into the camera. If the filter is placed directly in front of the camera lens, the shooting angle is generally within 20°.

Precautions for Large-Angle Interference Light

Even if a wide-angle field of view is not required, if there is large-angle interference light, especially if there is interference light with a slightly shorter wavelength than the signal light, it is also recommended to place the filter inside the camera, close to the CCD or CMOS and behind the lens. This is beneficial for reducing interference light in the wide-angle field of view.

Elimination of Double Image Interference

For applications such as size recognition and face recognition, where high-quality initial images are required, conventional filters have a relatively high reflectance for passband wavelengths. For built-in ultra-thin filters, the two surfaces are generally coated, and the front and rear glass surfaces work together to complete the wide-cutoff task, which means that both surfaces of the glass substrate are coated with complex multilayers with cutoff function. In this way, the two surfaces in the passband are prone to generate relatively large residual reflections. For example, if the passband transmittance of the front surface is 95% and the passband transmittance of the rear surface is also 95%, the two surfaces have 5% residual reflection each, and multiple reflections will occur between the two surfaces. This phenomenon will cause double images of the CCD or CMOS image, and the sharpness of the image contour will be reduced.