Facial Structure Detection Using Neural Networks

Facial Recognition or diagnosis as very well-known debatable topic nowadays as it has advanced to great lengths in recent years. The pivotal reason for this breakthrough is increase in computational power, while this has proved to break many stereotypes for machine learning algorithms but it has also presented a challenge for smaller scale devices. While deep learning based techniques for object detection have enhanced quickly over the recent years, most ways to deal with face detection are as yet dependent on the CNN framework, prompting constrained accuracy and execution speed. The scope for this project is largely based on making an efficient algorithm that is able to process on a smaller scale system by reducing the features used to detect algorithms and maintaining the accuracy at the same time. Reducing nodal points in creating biometric map of a face with best accuracy of the algorithm. The main goal is to implement a CPU based Facial diagnosis system by using the core python libraries and scholarly articles and also improvising on the speed and accuracy of the same. Reduced the frame rate and display it in low resolution for faster execution of trained models.

Face detection is a vital considered problem in Artificial Intelligence technology. Current technology detectors can without much of a stretch identify close front faces. Recent research here to emphasis more on the uncontrolled face recognition, where various aspects, for instance, and remarkable insight can produce large visual dissimilarity in face appearance, and can extremely devalue the power of the face detector. Face is a unique attribute of humans. The facial structure of ‘identical twins’ also differ in some aspects. As a normal human, we identify and distinguish among faces with very easy and unconcerned view. It was presumed that it would be very easy to detect forgeries.

The uniqueness of human faces is also a strong reason for their extensive use in applications where recognition or identification of people is important. At the beginning, we must discern between face recognition and person identification. In facial recognition only the visual inducement is used. On the other hand, In the case of person identification, however, other indications also play a vital role. From our research we understood in face recognition for the most part originate from two aspects: 1. The large visual dissimilarity of faces in the disordered foundations; 2. The extensive search space for possible facial positions and face dimensions. It is possible to implement more improved features in a practical face identification solution until the false alarm detections can be refused quickly in the early stages. In Comparison with the earlier homespun features, CNN can automatically learn features to capture complicated visual fluctuations by leveraging a huge amount of training data and its testing phase can be easily parallelized on GPU cores for expedition. In this project, we have decided to apply the Convolutional Neural Network (CNN) to face recognition.

 Importance of Facial Recognition

The Face Recognition (FR) is a vital research area because of the extensive applications in the fields of commercial and enforcement of law. A computer to perfectly recognise human faces would be crucial in several applications. This can be used for identification of criminals, as it’s been done in various ways which is not so efficient. As we have advancement in researched, the process can be made much time efficient and more robust since it will have only a minimum human interference. FR can also be used for authentication in security systems. Examples of such systems include computer systems and ATMs. In other words, all facilities that allow access based on the recognition and identification of the person can be embedded with this technology. Even with our ability to identify faces rather elegantly, is difficult to narrate a human face. The most common method is to specify different facial features, e. g. the hair is black, the eyes are brown, the nose is flat. The features play a vital role in the recognition process. Here important features are eyes, nose, lips, ears and hair. The dimensional relationships of the features are also important. However, the face is not just a sum or a collection of the internal features. The identification of faces involves unification of the facial features.

Various Techniques of Facial Recognition

Most facial recognition computer programs are based on locating and measuring selected features on face which are then compared with given corresponding measurements of known faces. Usual examples of the types of measurements used are ratios of distances between eye corners and mouth; measured ratio of height of face to width of face; and lines and angles of points along the face. Even though these processes resemble the sequential face recognition techniques used by people while performing complex face recognition tasks, they do not appear to be related to the immediate processes that people use when identifying familiar faces.

Conventional FR methods based on Visible Spectrum (VS) are facing challenges like object illumination, pose variation, expression changes, and facial disguises. As an enhancement, the Infrared Spectrum (IRS) can be implemented in human FR. The IR based Multi/Hyperspectral Imaging System can decrease the several constraints induced in the existing FR systems. A different research approach, researchers have used feature based techniques to produce the features of IR images for FR. Basically Features of an image are extracted based on Local Binary Patterns (LBP), Wavelet Transform, Curvelet Transform, Vascular Network and blood perfusion. The wavelet transform is applied to represent 1-D and 2-D signals, including face appearance. The Curvelet transform enhances the functioning of wavelet transform in which degree of orientation localization directly relay on the Curvelet.

 Facial Detection Using Convolutional Neural Networks

In 1994 Valliant implemented neural networks for face recognition. In their work, they proposed to train convolutional neural network to identify the presence or absence of a face in an image. In 1996, Rowley presented a retinal connected neural network for straight front face detection. The method was enhanced for rotation invariant face detection later in 1998 with a network to estimate the orientation and apply the proper detector. In 2002 Garcia proposed a neural network to recognise semi-front facing human faces in complex images; in 2005 Osadchy trained a convolutional network for simultaneous face detection and pose calculation.

One of the recent facial recognition based on CNN method is the R-CNN by Girshick which has achieved the state of the-art result on VOC 2012. R-CNN follows the “recognition using regions” pattern. This generates category of non-dependent region proposals and pulls out the CNN features from these non-dependent regions. Further it applies class-specific classifiers to identify the object category. In Comparison with the general object detection task, unsupervised face identification presents different tasks that make it impractical to directly apply the R-CNN method to facial identification. Generally, the problem exists in object/facial detection. This has to explicitly addressed this with CNNs. Rather than training a CNN for bounding boxes regression as in R-CNN, the training of a multi-class classification CNN for analysis is carried out. From this one can observe that a multi-class calibration CNN can be smoothly trained from limited amount of training data, while a regression CNN for measurement anticipates more training data. It is believed that the discretization decreases the difficulty of the measurement problem so that it can achieve good measurement accuracy with simpler CNN structures.

In comparison with these face detection systems, one of the work learns the classifier directly from the image instead of depending on homespun features. Hence this benefits from the powerful features learned by the CNN to better recognise faces from highly cluttered backdrops. At the same time, detector is much faster than the model-based and exemplar-based recognition software and has a frame rate comparable to the classical boosted cascade with normal features. Along the advantages of the CNN, the detector is easy to be embedded on GPU for much faster detection and efficiency.

As growing popularity of Facial Recognition in a day-to-day basis the significance of the study can be highly anticipated as approaching to more scalable solutions for smaller systems without comprising the core implementation idea of the algorithms. By the end of the implementation of our core concepts we should expect the system to perform accurately on smaller scale devices without high processing powers and our goal is to reach the same output achieved on a higher end system. In advancement of this we can try to implement the same techniques for object recognition and in the end combine both approaches to create a more scalable system.