PKT FOR FAST AND EFFICIENT HUMAN SKIN DETECTION
by DR GRACE L. SAMSON
ABOUT PKT
PKT (packed k-dimensional tree) is new learning algorithm, which improves the performance of learning procedures in terms of
speed, accuracy, and efficiency, using spatial inference, and algorithm. It
is a new rule-based fast algorithm that depends on an improved spatial structure
to perform object detection in computer vision and pattern recognition.
We applied the new model to human skin detection from an image; however, the method can also work for other machine learning tasks involving image pixels.
In this project, we proposed an improved RGB/HSL human skin color threshold to tackle darker human skin color detection problem. We combined the new color model with PKT for human skin/face detection from colored 2D images. We also implemented a novel packed quad-tree (PQT) to speed up the quad-tree performance in terms of indexing.
We compared the proposed system to traditional pixel-by- pixel (PBP)/pixel-wise (PW) operation, and quadtree based procedures. The results show that our proposed spatial structure performs better (with a very low false hit rate, very high precision, and accuracy rate) than most state-of-the-art models.
The unique thing about PKT is that only ≅20% of the elements of the dataset are required to make valid prediction/decision. This means that for an image, you will need just about 20% of the total pixels available, to classify the all the pixels accurately.
DIFFERENCE BETWEEN PKT AND SUPERPIXELS
The super-pixel (Sp) technique is different from the procedure of the PKT. Unlike the supervised/semi-supervised construction process (using an initial human marked image segmentation) of the Sp-based systems, PKT is fully automatic and unsupervised. It does not require hand annotated images, training sets, initial estimated regional clusters, or an input parameter (like ¢).
The steps in creating Sp(s) using the common SLIC algorithm, include:
1. Create initial regions based on ¢ (the parameter that determines the number of Sp(s))
2. Group the regions to cluster pixels using a similarity criterion.
3. Modify the connection between regions.
Thus, based on ¢, the number of pixels in a Sp would be:
𝑌 = √𝑁⁄¢ …. 1
Where 𝑁 is the number of pixels in the input image, ¢ is the estimated number of super-pixels (sub regions)
constructed, and 𝑌 is the number of pixels in each Sp(s).
Rather than find 𝑌, PKT computes Δ as in the equation below
Δ = ⌈((DL)/ μ) e⌉…. 2
In the above equation, Δ is compared to ¢, DL to 𝑁, and 𝑌 to μ. e is 1 all over the the number of dimensions or attributes (k) in the dataset
Note: while ¢ is estimated in Sp(s), Δ is automatically computed in PKT based on k, the number of dimensions of the image or the number of principal attributes of the data table (for non-spatial data).
PKT is a multidimensional structure and can apply to higher dimensional space or high dimensional datasets. Moreover, the projected number of regions of Sp(s) based on ¢, can lead to over-segmentation, whereas the number of PKT sub-regions Δ is optimally computed based on the fixed determinant value (μ = 25).
In addition, setting an initial position for Sp regions using constant distance increments, presents the challenge of possibly placing these positions (centres) at an image border and thus, Sp might fail to obtain a good segmentation. This is not the case with PKT.
PKT cells (regions) are generated automatically and instantly through a recursive partitioning strategy. Furthermore, only 5 strategic pixels are selected from each PKT sub-region, as against computing the sum of the probabilities of pixels of Sp’s sub-regions.
Finally, the use of predefined determinants μ to automatically detect the optimal number of PKT sub-regions, reduces the time taken to find the skin colour pigments, by eliminating the daunting search for ¢.
Comments
Post a Comment