This new technology encourages researchers in sign language recognition to customize it in real-time recognition. Integrating Speech and Hearing Challenge Individuals at omain features (Fourier transform) plus some information from motion analysis for ecognising 20 words. The data set was collected from 20 signers but the system is person ependent. HMMs were used as the classifier. An average of 92.5% recognition rate was chieved. Zieren and Kraiss (2004, 2005) presented two systems for isolated recognition: he first is for recognising GSL, on a vocabulary of 152 signs achieving a rate of 97%, ising HMM, but the rate decreases for the group of signs that contain overlaps in either ands or face and hands. Compared to other sign languages, not much has been done in he automation of the Arabic sign language, except few individual attempts (Mohandes 001; Jarrah and Halawani 2001). Al-Rousan, Aljarrah, and Hussain (2001) and Jarrah nd Halawani (2001), respectively, developed two systems for recognizing 30 static estures of Arabic sign language using a collection of Adaptive Neuro-Fuzzy Inference system (ANFIS) networks for training and classification depending on spatial domain eatures. Assaleh and Al-Rousan (2005) used coloured gloves for collecting a varying size lata samples for 30 manual alphabet of Arabic sign language. Polynomial classifiers were ised as a new approach for classification. In a recent (2005) work in Arabic Sign sanguage, Mohandes, and Deriche (2005) developed a system that recognised 50 signs of vords performed by one person having 10 samples per sign. They achieved 92% ecognition accuracy. Tolba, Abdellwahab, Abulle-Ela, and Samir (2010) used Pulse- oupled neural network (PCNN) for recognition of 30 alphabets postures and gained 93% ecognition accuracy. The system is signer-independent and achieved system invariance gainst rotation, scaling and colour. Earlier, researches in sign language recognition uffered from two main problems. First, the recognition accuracy is sensitive to any hanges like the signer, background or geometrical transformations. Second, all the xisting models find difficulties in identifying some postures. This is caused by the fact that . single view for the hand does not distinguish between two different postures (Figure 1). Tolba et al. (2010) overcame the first problem by customising PCNN to generate image ignature. This signature represents transformation-invariant image features. He failed to Ey Ae mre are of Beene geehaso: ce vetawswl= «soma ‘SET bie BP tse ce i PR See! Be Figure 1. Two different postures (a) front view and (b) side view. Figure 2. (a) A modified model for PCNN neuron and (b) receptive field in (a). There are several differences between the algorithms for the modified PCNN neuron and the exact physiological pulse coupled neuron. The differences are due to several simplifications made to the calculations, while still keeping the main features of the general theory. Each neuron in the modified PCNN could be described by the following set of equations (Kinser and Nguyen 2000): Table 1. Examples of previous 3D recognition systems. Tangruamsub, Takada, and Hasegawa (2011) presented a 3D object recognition method. The proposed objectives are to overcome shortcoming of the appearance-based method, which lacks a spatial relationship between the parts of an object, and those of other 3D model methods, which require complicated computation. The proposed method s based on a voting process. Appearance estimation is introduced in this work in order to Jeal with the faulty detection problem. He tested the method for object detection and pose >stimation. Figure 4. The hands 3D model construction. training data and nine image sets for testing data. The acquired images at (0°, 90°), (10°, 100°),... and (90°, 180°) were used as the training set and the rest of the images (images at (5°,95°), (15°, 105°),... and (85°, 175°)) were used as the testing set. The training data set is used to build the 3D object model in the recognition stage. Figure 4 illustrates the 3D feature generation operation. TN Ne a i ee eg ee eee es ee ee el The 2D-captured images are then digitised and sent to the PCNN Feature Generation module. The two output signatures are then weighted and combined linearly to produce Table 2. Value of wl and w2 with different capturing angles. 3D image signature. The PCNN 2D feature generation module uses (15) as image signature generation technique. Let P1(i) represents the generated signature at point A and P2(i) at point B, one defines F(z) as 3D image signature. It is computed as the weighted linear combination of them: Figure 5. The signer and capturing cameras positions. Figure 8. The corresponding NFA for counting from | to 2 gesture. Figure 9. The number of postures (1 hand and 2 hands) against the number of gestures. Table 3. The effect of wl and w2 on recognition accuracy. Figure 10. The recognition accuracy against the object percentage. Figure 11. ROC curve of the three proposed weights combinations. Table 4. The gestures of the dataset. Figure 12. Recognition accuracy against the number of gestures. Table 5. Comparative study with existing systems.
