- Количество слайдов: 33
Face Detection Using Neural Network By Kamaljeet Verma (05305905) Akshay Ukey (05305045)
Problem Definition Identify and locate human faces in an image regardless of their: – Position – Scale – Orientation – Illumination
Motivation A challenging problem faced by Computer. Vision community. Face is a highly non-rigid object. Main step before Face Recognition. First step in many surveillance systems. Example applications: – Automated security systems – Intelligence information – Robotics
Face Pattern Space Consider 19 x 19 thumbnail face pattern. Possible combinations of gray values is 256361 = 22888 Number of galaxies in the Universe ≈ 235 Extremely high dimensional space.
Difficulties in Face Detection Facial Expressions – Smiling, frowning, etc. Presence or absence of structural components – Beard, moustache, glasses, etc. Pose – Frontal, upside down Position – Location of the face in the image.
Difficulties in Face Detection Scale – Size of the face in the image can vary. Orientation – Face appearance directly vary for different rotations about the camera’s optical axis Illumination – Images taken in different lighting conditions adds to the variation in face pictures.
Example Images from CMU dataset
Current Research Representation of a typical face in the computer. Search Strategy Increasing the speed of the process of system. Achieving accuracy. Combining detection results.
Appearance based method Models – Learned from a set of training images. Training set – Captures the representative variability of facial appearance.
Build Training Set Proper mix of positive and negative examples. Positive Examples – Having as much variation as possible. – Manually resize each into a standard size(e. g. : 19 x 19) Negative examples – – – Images not containing face Large image subspace Bootstrapping
Representation Raster scanned image represented by a vector of intensity values. Block-based – Process each image as if divided into blocks. – Blocks may be overlapping or nonoverlapping
Pre-processing Masking – Minimize background noise in face image Illumination Gradient Correction – To minimize heavy shadows due to lighting angles. Histogram Equalization – To compensate for difference in illumination brightness, skin colours, camera responses, etc.
Masking l Remove near-boundary pixels with 19 x 19 binary mask. l For avoiding unwanted background structure from face image. l Effectively reduces the 19 x 19 pixel window vector space.
Illumination Correction Take an image of 21 x 21 pixels. Divide it into 7 x 7 blocks. For each block compute minimum intensity pixel giving 3 x 3 minimal brightness plane. Resize the plane to 21 x 21. Subtract this plane from original image.
Histogram Equalization l. Equalize intensity values. l. Expand range of intensities in the window.
Classifiers Different classifiers can be used. Classifiers – Neural Network – Principal Component Analysis – Support Vector Machines – Naives Bayes Classifier
Neural Network Approach
Neural Network - Introduction A modeling technique – Based on the observed behavior of biological neurons. – Used to mimic the functioning of brain. Features – Ability to adapt to new environments. – Made up of large number of processing units. – High processing speed. – Used to solve many complex problems.
Components of a Neural Network Four Main Components – Processing Units (pj) Each pj has a certain activation level – Weighted Interconnections Determine how the activation of one unit leads to input for another unit. – An activation rule Used to produce output signals. – A learning rule Specifies how to adjust the weights for a given input/output pair.
The Perceptron Model A perceptron is a computing element with input lines having associated weights and the cell having a threshold value. Model motivated by the biological neuron. Output (y) Threshold (θ) w 1 x 1 w 2 x 2 . . . wn x 3 Weights Inputs
NN in Face Detection? Neural Nets can classify data into a given set of classes. Face Detection Classes – – Face class Non Face class Face Detection Input and Output – The shade of GRAY of each pixel is presented to the neuron in parallel. – E. g. for a 10 X 10 pixel image, there will be 100 input lines x 1 to x 100, with respective weights w 1 to w 100. – The output y will represent the presence or absence of a face.
NN based Face Detector (Rowley) Two Stages: – It applies a neural network-based filter to an image. – It arbitrates the filter outputs. Filter – examines image at several scales. – Detects locations containing faces. Arbitrator – Merges detections from individual filters – Eliminates overlapping detections.
Stage 1. NN based Filter 20 X 20 window is extracted from input image. Preprocessing – – – Applied to the 20 X 20 window. Attempts to equalize the intensity values across the values. Steps • • Illumination Correction Histogram Equalization
Stage One Neural Network – Input is the 20 x 20 preprocessed window. Input Layer – Consists of 400 pixel intensity values. Hidden Layer – – Consists of 3 types of hidden neurons 4 which look at 10 x 10 pixel subregions 16 which look at 5 x 5 pixel subregions 6 which look at overlapping 20 x 5 pixel subregions. Output Layer – Single neuron having real value in the range [-1, 1] – Indicates if the window contains a face or not.
Algorithm for Face Detection
Stage One Training – Neural network is trained using standard backpropagation algorithm. – Done on face examples gathered from face databases at CMU and Harvard. – Face examples easy to find. – Non face examples Very large space Collecting small “representative” set is difficult. – Bootstrapping technique can be used.
Stage One Bootstrapping 1. Start with a set of non-face examples in the training set 2. Train the neural network with the current training set. 3. Run the learned face detector on a sequence of random images. 4. Collect all the non-face patterns wrongly classified as faces. 5. Add these non-face patterns to the training set. 6. Go to step 2 or stop if satisfied.
Stage One Scaling Scan an input image and run the algorithm. Scale down the image by a factor of 1. 2 and run the algorithm again. (Process continued until image size is too small).
Stage Two: Arbitration Merging overlapping detections within one network. Use multiple networks – initialize them to different initial weights. – Run the algorithm. – Different sets of negative examples will result. – Arbitrate among their outputs. Eg. Signal face detection only when all the networks agree that there is a face.
Experimental Results False Detect False Miss The algorithm can detect between 78. 9% and 90. 5% of faces in a set of 130 test images, with an acceptable number of false detections.
Conclusion The neural network approach has better performance in comparison to other approaches. The technique reduces the processing time by not using fully connected network. The approach is heavily dependent on the training set selected.
References H. Rowley, S. Baluja, and T. Kanade, Neural Networkbased Face Detection, Proc. of IEEE Conf. on CVPR, 1996. H. Rowley, S. Baluja, and T. Kanade, Neural Networkbased Face Detection. IEEE Trans. Pattern Anal. Mach. Intelligence, 1998. K. K Sung, and T Poggio. Example-based learning for view-based human face detection. IEEE Trans. Pattern Anal. Mach. Intelligence, 1998. Recent Advances in Face Detection, IEEE ICPR 2004 Tutorial, Cambridge, United Kingdom, August 22, 2004.