Digital Image Processing Face Detection Shrenik Lad Instructor: Dr. Jayanthi Sivaswamy

Transcription

1 Digital Image Processing Face Detection Shrenik Lad Instructor: Dr. Jayanthi Sivaswamy Problem Statement: To detect distinct face regions from the input images. Input Images:

2 Algorithm: Convert the input image into YCbCr space. YcbCr space segments the image into a luminousity component and color components. We try to remove as much non-skin color pixels as possible. The Cr and Cb components give a good indication whether a pixel is part of the skin or not. The threshold that are chosen are 150 < Cr < < Cb < 117 We make a binary image in which skin-colored pixels are white whereas all others are zero. Erode the binary image with a disk shaped structuring element of radius 6 and then dilate the resulting image with a disk shaped structuring element of radius 10 Fill the gaps in the image by hole-filling algorithm. At this stage, the areas in the image whose size is less than pixels are removed. The image contains areas representing face and non-face (hands, legs and background). After this we do Standard deviation thresholding on the image. Generally, face regions have higher std. deviation than non-face regions(hands, legs), because of the presence of eyes, nose etc. We do this by first finding connected components in the image. After that, we find the std. deviation for each component and decide a threshold for removing non-face regions. The resulting image should have all face regions highlighted in it.

3 Input Image Convert to YcbCr space Ycbcr image Skin colored pixels - thresholding only skin-colored pixels retained Binarisation Erosion and Dilation - Opening small areas in the background are removed Std. Deviation Thresholding on Connected components Non-facial regions removed Image with all Face regions

4 Input image 1 step wise results Input Image Ycbcr thresholding 3 faces to be detected in image Skin-colored pixels in image. Background areas also present Binarised image Erosion Binary image having 2 values, white for skin White region shrinks and small areas in colored. Otherwise black background removed

5 Dilation and hole filling Colored image White region expands in image only face regions and some non-face regions (hands, legs) present Connected Components Std deviation thresholding Different connected components in image Non-facial components removed by thresholding

6 Final Output Image: Observations : (step wise observations are written with results) The algorithm worked well for image1. The three faces are properly detected and are coloured with green in the input image. Thresholding in Ycbcr space seems to be very important steps. We should ensure that the skin-colored areas in background must be small in size, so that they can be removed by opening operation. The parameters used are chosen such that the algorithm gives best results for image 1.

7 Same parameters applied on Image 2: Output: Observations: 3 out of 4 face regions are detected but not completely Non-facial regions like hands of second child is wrongly detected The face region of third child is lost during the Ycbcr thresholding step itself, maybe because it does not lie in the skin-color range used in image 1 Because of shadow on the face of fourth child, only the forhead is being detected.

8 Same parameters applied on Image 3: Output: Observations: None of the face regions are being detected from the image Some of the loss is taking place during Ycbcr thresholding and then during erosion phase, all face regions are being lost. This might be because of the size of kernel used for erosion. The parameters need to be changed for good results on this image

9 Input image 2 step wise results Input Image Ycbcr thresholding 4 faces to be detected in image Skin-colored pixels in image. Background areas also present Binarised image Erosion Binary image having 2 values, white for skin White region shrinks and small areas in colored. Otherwise black background removed

10 Dilation and hole filling Colored image White region expands in image only face regions and some non-face regions (hands, legs) present Connected Components Std deviation thresholding Different connected components in image Non-facial components removed by thresholding

11 Final Output Image: Observations : (step wise observations are written with results) The algorithm worked well for image2 with different parameters. Except second child, 3 other faces are properly detected and are coloured with green in the input image. The face of second child is not completely detected. It got lost during the Std. Deviation thresholding, because its std deviation was less than the threshold. If we increase the threshold, the hand of second child is wrongly detected as face. (see the next image) Also, the face detection on second child is not uniform because shadow is present on the face. For others, the entire face is uniformly detected. A part of the hut in the background is also wrongly detected as face. It was not removed during erosion phase, because its size was more than the kernel mask. Kernel size could not be increased more, as it leads to loosing face components The parameters used are chosen such that the algorithm gives best results for image 2.

12 Output with different threshold Here, a different threshold was used during std deviation thresholding. But unfortunately, the hand-component of second child also satisfied the condition, and it is wrongly detected as face. Rest, everything is similar as in the previous output.

13 Input image 3 step wise results Input Image Ycbcr thresholding many faces to be detected in image Skin-colored pixels in image. Background areas also present Binarised image Erosion Binary image having 2 values, white for skin White region shrinks and small areas in background removed colored. Otherwise black

14 Dilation and hole filling Colored image White region expands in image only face regions and some non-face regions (hands, legs) present Connected Components Std deviation thresholding Different connected components in image Non-facial components removed by thresholding

15 Final Output Image: Observations : (step wise observations are written with results) The algorithm does not work so well for image3. Along with the faces, hands of 3 people are being wrongly detected. This is because of the std deviation threshold. The face of one person at the back is not at all detected. Some part of its face got lost during erosion-dilation phase, and the remaining part was lost in std deviation thresholding. On changing the kernel size used for erosion-dilation, less faces are detected. Also, decreasing threshold more gives poor performance. Some part on the floor could not be removed during erosion because its area was little more than the kernel size. Kernel size could not be changed because face components were lost. The parameters used are chosen such that the algorithm gives best results for image 3.