Image Signature Improving by PCNN for Arabic Sign Language Recognition M. Fahmy Tolba1, M. Saied Abdel-Wahab 2, Magdy Aboul-Ela3, and Ahmed Samir4 1,2,4 Faculty of Computers and Information, Ain-Shams University; Cairo, Egypt. 3 Sadat Academy for Management Sciences; Maadi, Cairo, Egypt.

Abstract This paper offers the problems of standard image signature generation and their standardizations for image recognition using Pulse Coupled Neural Network (PCNN). The aim of this research is to propose new technique for image signature using PCNN. This new technique is used for Arabic sign Language (ASL) static alphabets recognition. The new model mainly adds the continuity factor as a weight of the current pulse in signature generation process. This modification preserves the invariance property and enhances the feature selection for image recognition purposes.

Keywords Arabic Sign Language (ASL) Recognition, PCNN, Discrete Fourier Transform (DFT), Multi-layer perceptron (MLP), continuity Factor.

1- Introduction In 1990,Mammalian neuron model was obtained by Eckhorn et al [1]. through analyzing the oscillation phenomenon of synchronous pulse bursts in the cat visual cortex[1]. In 2000, a new neuron model was proposed by Kilter and Leo [12] to analyse the correlation of a couple of adjacent neuronal ignition quality. In the same year, Bressloff and Coombes[14] made an intensive study of the dynamic behavior of PCNN. Different PCNN algorithms were developed to simulate the exact model [20]. These modifications are developed in purpose of the dimension decreasing of image classification space and for image recognition. They are focused on optimization of standard PCNN mathematical model. PCNN has been widely applied to image denoising[4], image smoothing[8], image processing [8], image segmentation[4], image fusion[11]. PCNN can transform 2D images into 1D time signal with some interesting features [3]. Raul[15] used these time series for pattern recognition. He proposed the standard approach of feature generation using Modified PCNN based on the sum of output quantities of all neurons for n iterations. Mokriš[20] proposed a new form of features generation which is based on features standardization. This paper proposes a new approach which emphasizes the standardization concept. This approach adds a weight factor for each iteration. This approach was implemented and tested with the previous approaches and applied in Arabic Sign Language (ASL) Recognition. Section 2 and 3 in this paper explain different modifications and algorithms for PCNN models. Section 4 discusses in details the previous feature extraction methods. Section 5 explains the recognition model used for ASL recognition, and how the new approach was used as part of it. Section 6 illustrates the experimental results. Finally section 7 states the conclusion of this work.

2- Modified PCNN 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 simplify 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 [2][11]: L(i) = L(i-1) · e (-αL) + VL · (R*Ysur (i-1)) F(i) = S + F(i-1) · e(-αF) + VF · (R*Ysur (i-1)) U(i) = F(i) · [1 + β· L(i)]. θ(i) = θ(i - 1)e-(αq) + Vθ Yout(i-1) U > θ(i) => Yout = 1 (Firing Condition) otherwise => Yout = 0 Where S is pixel input intensity.

(1) (2) (3) (4) (5)

Ysur is the firing information whether the surrounding neurons have fired or not, L is the linking, F is the feeding, θ is threshold function and Yout is whether this neuron fires or not. An example of the modified PCNN neuron architecture is shown in Fig1 as a schematic block diagram of the modified PCNN neuron as described by equations (1) - (5).

Fig 1 Original Modified Neuron

3-Optimized PCNN The main aim of Optimized PCNN was to reduce the number of generated features to reach high image recognition performance. The optimization was based on the PCNN with modified primary input (MPCNN), where the following disadvantages were eliminated [11]: 1. high number of parameters and problems with their optimization, 2. optimal number of iteration steps determination, 3. value of the most significant feature equals to 1. after standardization so this feature loses its information value. The feeding potential F(i) is defined by the intensity pixel Sij only as in case of Modified-PCNN. The linking potential L(i) is defined only by the convolution matrix K(i) that is calculated by term K(i)= R*( Xsur (i-1).Ysur (i-1)). L(i)=K(i). U(i) = F(i) · [1 + β· L(i)]. X(i)= X(i)> 0.5 => Yout = 1

(6) (7) (8) (9)

(Firing Condition) (10)

otherwise => Yout = 0 where Y(i) is output quantity based on step-function and X(i) is output quantity based on sigmoid function.

4- Feature Generation Methods The standard approach of feature generation by MPCNN is based on series of virtual binary images generation. These binary images are produced by activated neurons (Yi = 1) and non activated neurons (Yi =0) for each iteration step(i). The value of feature G(n) for specific iteration (i) is calculated as sum of output quantities Yij of activated neurons in the given iteration step [15]. G(n)=

(11)

Through influence of geometrical transforms it is very important to standardization of the generated features by standard equation [20].

g(n)=

(12)

The feature with maximal value represents the iteration step (i), in which the most neurons were activated at the same time. The virtual binary image in the given iteration step comprises the best segmentation result of input image. Therefore it is possible to consider this feature as feature with maximal information value for image recognition process. But, using equation (12) concludes that the value of this feature will be always 1. This makes this feature irreverent for image recognition process. To solve this problem, a modification can be applied to equation (12), such that g(n)= where 0≤

(13) is the intensity of a given image pixel (i,j), such that: ≤1

(14)

The left side of condition (14) is always valid. The right side of condition (14) is not valid in the all cases, because the sum of values Yi in the given iteration step I may be higher than sum of values Sij. Mokriš [15] proposed a new form of feature generation. He introduced a new equation for feature value calculation g(n): g(n)=

(15)

He also proved the validation condition 0.5