Using LabVIEW and Vision Development Module to Create a Robust Automated Face Detection System

  Print

"We created a system capable of detecting multiple faces using the LabVIEW graphical development environment and Vision Development Module to perform complex, tedious tasks. NI products eased the development process with built-in functions in the basic package as well as complex imaging functions in the NI Vision Development Module."

- Alex See, Monash University Malaysia

The Challenge:
Creating a reliable face detection system that can accurately detect faces in a cluttered image.

The Solution:
Using NI LabVIEW software and Vision Development Module to develop a fast, automatic face detection system that can run on any standard computer system without additional hardware requirements.

Author(s):
Alex See - Monash University Malaysia
K.Y. Tan - Monash University Malaysia

Accurate face detection in a video feed has many uses, including facial recognition, closed-circuit television (CCTV) monitoring, and as an autofocusing feature for video conferencing or web cameras. However, the vast image variations in a cluttered and uncontrolled environment require substantial processing to accurately detect faces with acceptable speed.

Human face detection in a cluttered environment is a notorious problem for researchers in the computer vision community. The face detection system we created aims to properly locate the region(s) in the video image consisting of a human face. The first step is effectively reducing the region required for computation. We achieve this with segmentation by color and background subtraction.

To properly locate an exact, relevant, and appropriate image of the face, the system accurately matches an average face template using normalized cross-correlation. We perform the correlation on relative gradient images, a novel image processing method suggested by Wei and Lai, offering lighting invariant images. This method overcomes the limitation of the grey-level image template matching in lighting variations.

Color Segmentation

To segment the skin-colored regions, we use hue-saturation value (HSV) because the skin tones of different ethnic groups actually have a similar HSV, with only the illumination differing. Hue-saturation plane is suppose to be unaffected by differences in illumination caused by changing lighting conditions.

We use a combination of built-in morphological functions in the Vision Development Module to carry out color segmentation (see Figure 3). We segment the skin-toned pixels using threshold for the hue and saturation values. The particle filter removes any small, noisy particles that dilate as the function closes holes to condition the large particles for easier processing later.

The hue-saturation plane is supposedly not affected by illumination differences, however, we found that this is true only to a certain limit of brightness or darkness. To achieve acceptable system performance, we must segment skin-like regions in various lighting conditions. Note that natural sunlight has differing colors throughout the course of the day and artificial lighting differs in color planes other than illumination.

Background Subtraction

To meet these varied requirements, the threshold limits should allow for a wider range of hue values, but this results in a large occurrence of error detection of skin-like hue regions in the image. For this reason, we use background subtraction to aid in the segmentation. By performing background subtraction, we can rule out any skin-tone-colored items in the background.

Background subtraction uses a combination of basic operators to add, subtract, and divide the pixel values between the background hue and the video frame hue. The hue plane offers lighting invariance while the saturation plane is not considered to simplify the processing.

Figure 4 shows the results in which intermediate grey values are the actual matching pixels (due to the mathematical manipulation), and the lighter and the darker the deviation from the grey, the larger the difference between the pixel values. This image is transformed to binary using multiple thresholding and later processed using particle filter and dilation before combining it with the color segmentation result.

Combined Segmentation

After combining the two segmentation results, we remove more particles and perform particle analysis to use the bounding boxes of the segmented particles for face matching.

Face Matching

Once the color segmentation and the background subtraction results merge, the program uses the information for face matching. Face matching is necessary because the segmentation result consists of any skin-colored nonbackground items in the video frames, including garments, hands, and neck. The system must accurately locate the face to obtain proper images containing the facial features with the optimal region of interest.

The grey-level image template matching has high failure rates under lighting conditions that differ from the template image. To counter this problem, we use the relative gradient image approach suggested by Shou-Der Wei and Shang-Hong Lai [1]. The relative image gradient is given by the following equation [1]:

http:/vcm.natinst.com:27110/cms/images/casestudies/dcuipsei5282178929124485004.jpg

Where I(x,y) is the image intensity function that denotes the gradient operator that takes the partial differentiation across the x and y direction, W(x,y) is a local window centered at location (x,y), and c is a constant (to avoid division by zero). This equation provides some “smoothing” effect to the gradient value due to the division with a local maximum resulting in a possible lighting invariant gradient image.

The relative image gradient of both the average face and the video frames detect the faces in the segmented regions. We use the relative gradient image of the average face as the template for template matching and perform it using normalized cross-correlation (the pattern matching functions in LabVIEW). We use the incremental pyramidal search for each of the segmented particle—the size of the images to check for matches start from small to large. We do this to accommodate possible face size variations in video frames.

For faster and more accurate template matching, we shrink and blur the images. We use small templates and shrink the images used to search for faces down to the size of the template and increase their size through the search. The results show that this approach is effective and even detects faces with glasses on, at different angles, and at varying distances.

After detecting the faces, they are extracted and reoriented to become straight, using the results from the successful pattern matching.

Conclusion

We created a system capable of detecting multiple faces using the LabVIEW graphical development environment and Vision Development Module to perform complex, tedious tasks. NI products eased the development process with built-in functions in the basic package as well as complex imaging functions in the NI Vision Development Module. This system is useful in various surveillance activities. With the minimal use of expensive industrial grade hardware, such as a CCD camera, we created a low–cost but robust face detection system.

Author Information:
Alex See
Monash University Malaysia
No 2 Jalan Kolej, Bandar Sunway
Selangor 46150
Malaysia
Tel: +60 56360600
Fax: +60 56329134
alex.see@eng.monash.edu.my

Bookmark and Share


Explore the NI Developer Community

Discover and collaborate on the latest example code and tutorials with a worldwide community of engineers and scientists.

‌Check‌ out‌ the‌ NI‌ Community


Who is National Instruments?

National Instruments provides a graphical system design platform for test, control, and embedded design applications that is transforming the way engineers and scientists design, prototype, and deploy systems.

‌Learn‌ more‌ about‌ NI