PAPER
Adaptive Morse Code Recognition Using Support Vector Machines
for Persons with Physical Disabilities
Cheng-Hong YANG†, Member, Li-Yeh CHUANG††, Cheng-Huei YANG†††a),
and Ching-Hsing LUO††††, Nonmembers
SUMMARY In this paper, Morse code is selected as a communication adaptive device for persons whose hand coordination and dexterity are im-paired by such ailments as amyotrophic lateral sclerosis, multiple sclerosis, muscular dystrophy, and other severe handicaps. Morse code is composed of a series of dots, dashes, and space intervals, and each element is trans-mitted by sending a signal for a defined length of time. A suitable adaptive automatic recognition method is needed for persons with disabilities due to their difficulty in maintaining a stable typing rate. To overcome this prob-lem, the proposed method combines the support vector machines method with a variable degree variable step size LMS algorithm. The method is divided into five stages: tone recognition, space recognition, training pro-cess, adaptive processing, and character recognition. Statistical analyses demonstrated that the proposed method elicited a better recognition rate in comparison to alternative methods from the literature.
key words: Morse code, adaptive signal processing, least-mean-square algorithm, support vector machines
1. Introduction
Most computer or high technology related products are de-signed for able persons, and are inaccessible to persons with disabilities. A current trend in high technology production is to develop adaptive tools for persons with disabilities to as-sist them with self-learning and personal development, and lead more independent lives. Among the various technolog-ical adaptive tools available, many are based on the adapta-tion of computer hardware and software. The areas of appli-cation for computers and these tools include training, teach-ing, learnteach-ing, rehabilitation, communication, and adaptive design [1], [2].
Many researchers have focused on a reduced set of input keys since the unadapted computer keyboard is not a useful communication tool for persons with disabilities. Many computer assisted key-in systems, e.g., the head mouse, mini-keyboard, king-keyboard, trackball, joystick,
Manuscript received December 20, 2004. Manuscript revised December 31, 2005. Final manuscript received March 27, 2006.
†The author is with the Department of Electronic Eng.,
Na-tional Kaohsiung University of Applied Sciences, Kaohsiung, 807 Taiwan.
††The author is with the Department of Chemical Eng., I-Shou
University, Kaohsiung, 807 Taiwan.
†††The author is with the Department of Computer
Communi-cation Eng., National Kaohsiung Institute of Marine Technology, Kaohsiung, 807 Taiwan.
††††The author is with the Department of Electrical Eng., National
Cheng Kung University, Tainan, 701 Taiwan. a) E-mail: [email protected]
DOI: 10.1093/ietfec/e89–a.7.1995
alternative keyboard, keyguard, and touch screen, have been developed to overcome barriers [1], [2]. A set of switches for these input devices with an efficiency rate approaching one key press per selected character is ideal. The most com-monly applied adaptive tool, Morse code, has been shown to be valuable in assistive technology (AT), augmentative and alternative communication (AAC), rehabilitation, and edu-cation [3]–[8]. An interactive Morse code emulation system helps beginners or persons with disabilities to become ac-quainted with Morse code and to communicate with a com-puter through the implementation of a multimedia package [9]. An easy-to-operate input interface is used to facilitate access to the Internet [10]. Recently, an environmental con-trol system was designed to access electronic facilities [11]. Over 30 manufactures/developers of Morse code input hard-ware or softhard-ware for use in AAC have been identified to date [12].
A stable typing rate is strictly required to recognize Morse code characters. However, this restriction is a ma-jor hindrance, especially for persons with severe disabili-ties. Therefore, a suitable adaptive automatic recognition method of the keyed-in Morse code is needed. To recognize Morse code characters, two unstable elements, switch-down time (tone element) and switch-up time (space element), have to be predicted. Support vector machines (SVM) have been successfully applied to a number of real world prlems including recognition of handwritten digits, 3-D ob-jects, breast cancer prognosis, and engine-knock detection [13]–[16]. They demonstrate an impressive resistance to overfitting in classification and their training is performed by maximizing a convex functional, which means that there is a unique solution that can always be found in polynomial time. The proposed method, which combines the support vector machines method with the variable degree variable step least-mean-square algorithm [17], increases prediction accuracy. The results show a much better recognition rate when compared to previous results in the literature [4], [18]. This paper is organized as follows: in the next section, the new Morse code recognition method is presented, Sect. 3 de-tails the experimental result and discussion, with concluding remarks made in Sect. 4.
2. Method
Morse code is a system of asynchronous data bits com-posed of binary encoded circuit opposites (long-short) used Copyright c 2006 The Institute of Electronics, Information and Communication Engineers
for transmission and reception of alphanumeric information with which each character can be translated into a prede-fined sequence of dots and dashes (the elements of Morse code). A dot is represented as a period “.” while a dash is represented as a hyphen or minus sign “-.” Based on the def-inition of Morse code, the tone ratio (dot to dash) has to be 1:3. That means the duration of a dash is required to be three times that of a dot. In addition, the silent ratio (dot-space to character-space) also has to be 1:3 [18].
In 1996, Luo and Shih [3] proposed a system that could recognize varying typing speeds using an adaptive tech-nique, the Least-Mean-Square (LMS) algorithm. Their pro-posed method could adjust its characteristics to successfully recognize a message under unstable typing conditions, but the speed variation had to be limited to a range between 0.67 and 2.0 times the present typing speed. To satisfy this lim-itation, a user with disabilities had to be well-trained, oth-erwise the system would not successfully recognize his or her Morse code message. However, this limitation could not always be complied with by a beginner or those with very serious disabilities; and therefore, the method could not be effectively used. Subsequently, Shih and Luo [4] proposed an improved method that combined the Least-Mean-Square algorithm with a character-by-character matching technique to overcome this limitation. A combination of the variable degree variable step least-mean-square algorithm and learn-ing vector quantization (LVQ) method was also applied to the Morse code recognition problem [18]. In the present study we substituted the LVQ method with support vector machines, which improved the recognition rate compared to the previous method (LVQ) by about 3 percentage points.
The proposed method in this paper is divided into five stages: tone recognition, space recognition, learning pro-cess, adaptive processing, and character recognition. A block diagram of the Morse code recognition process is shown in Fig. 1. Initially, the input data stream is sent
in-Fig. 1 Block diagram of the Morse code recognition system.
dividually to either tone recognition or space recognition depending on down time (tone element) or switch-up time (space element). In tone recognition, the tone ele-ment value is first recognized as either a dot or a dash, and then sent to the learning process, which is used to recalcu-late the decision function. Simultaneously, in the tone buffer section, the recognized tone element (dot or dash) and each successive tone element are saved in a dot-dash buffer and a tone element buffer. Next, in space recognition stage, the space element value is recognized as being either a dot-dash space or a character space. If the space element value is rec-ognized as a character space, it is divided by 3.0 (due to the 1:3 ratio, which has to be observed in Morse code) before being fed into the adaptive processing stage. Otherwise, the space element value feeds directly into the adaptive process-ing stage. After a character space is obtained, the value(s) in the tone buffer is (are) sent to character recognition, which identifies this character [9].
A Morse code character, xi, is represented as follows:
e1(xi), b1(xi), . . . , ej(xi), bj(xi), . . . , en(xi), bn(xi)
1 j n where
ej(xi): jth tone duration in the character xi.
bj(xi): jth silent duration in the character xi.
n: the total number of Morse code elements in
charac-ter xi.
mj(xi): a dot or dash recognized from ej(xi).
sj(xi): a character space or dot-dash space recognized
from bj(xi).
2.1 Tone Recognition
Each tone element is treated by normalization to obtain an input value within a range of−1 to 1.
x = 2.0 ∗(tone element− 0.5 ∗ (tone max +tone min))
tone max−tone min
(1) where tone max and tone min are the largest and smallest values of the tone element respectively. If a tone element is larger than the tone max value, then the tone max value is substituted by this tone element value. If a tone element is smaller than the tone min value, then the tone min value is substituted by this tone element value. The obtained value x can be sent into a decision function to recognize the value as being either a dash ( f (x)≥ 0) or a dot ( f (x) < 0). The following equation applies: (2)
f (x)= sign ⎛ ⎜⎜⎜⎜⎜ ⎜⎝ i∈Is.v. αiyiK(xi, xj)+ b ⎞ ⎟⎟⎟⎟⎟ ⎟⎠ (2)
where αiare the optimized Lagrange multipliers. When
op-timized the Lagrange multiplier is called a support vector (αi 0). yi ∈ {−1, +1} and b are the bias [19]. The kernel
function used in this paper is a radial basis function (RBF), such as Gaussians Kxi, xj = e −xi−xj 2 2σ2 , i= j = 1, 2, · · · , (3) The new tone value of the input stream will be put into the decision function f (x) to determine the value as being either a dash ( f (x) ≥ 0) or a dot ( f (x) < 0). After tone recognition, the resulting value can be labeled and sent into the training data set. Then the training process is performed to recalculate the decision function.
At the beginning of this process the initial tone base (TB), which is used to serve as the initial dot-dash classi-fier, is absent. To determine the initial TB, the first nine values of tone elements are taken as reference values and sorted in descending order. Once the sorting is done, the ele-ment values are compared with each other to determine their relationship, and are either recognized as “dash” or “dot.” “Dash” means that a value is at least twice as large as any other value. A smaller value is defined as “dot.” After the “dash” or “dot” relationship is determined, the dash base and dot base represent the average of the dash values and the dot values. The resulting final values represent the ini-tial TB.
dot base= average of the dot values dash base= average of the dash values tone sum= dash base + dot base tone ratio= dash base / dot base TB= tone sum / tone ratio
For example, a Morse code digital stream is presented as follows: 203 496 675 470 195 1102 285 425 690 408 712 364 262 386 270 1087 244 1145, in which odd position data are defined as tones while even position data (underlined) are defined as spaces. The first nine tone values are thus 203, 675, 195, 285, 690, 712, 262, 270, and 244. After sorting, the descending order of these nine values is 712, 690, 675, 285, 270, 262, 244, 203, and 195. The first three
values (712, 690, 675) are more than twice as large as any other value in the sample and thus are dash values. Since the remaining values are less than half of any of the first three values, they are dot values. Thus, the sample data stream can be calculated as follows:
dot base= (285+270+262+244+203+195)/6 = 243.17 dash base= (712+ 690+ 675)/3 = 692.33
tone sum= 692.33+243.17 = 935.50 tone ratio= 692.33/243.17 = 2.85 TB= 935.50/2.85 = 328.25
As a result, the TB is 328.25, thus the training data set is {(203, −1), (675, +1), (195, −1), (285, −1), (690, +1), (712,+1), (262, −1), (270, −1), and (244, −1)}. Once the initial training data set is determined, it can be used to run the training procedure and find the initial decision function for the dot-dash classifier.
2.2 Space Recognition
The space recognition stage is employed to detect the spaces existing between whole characters, as well as the space between isolated Morse code elements, which comprise a unique character. Thus, if a data stream of characters com-posed of Morse code elements is entered, these elements must then be identified as being either spaces between whole characters or spaces between isolated elements of a charac-ter.
The procedure for this character detection operation is shown below:
1. initiate j=1.
2. if bj(xi) < silence base, then go to step 3, otherwise go
to step 4.
3. bj(xi) is a dot-dash space. Let j= j + 1 and go to step 2.
4. bj(xi) is a character space. Then a sequence of tone
du-rations between the character spaces is obtained. Go to step 1.
Unfortunately, the first character, xi, cannot be
im-mediately isolated because of the absence of an initial si-lence base (SB) value. Subsequently, the initial SB is ob-tained by extracting the first nine values of silent elements entered as reference values; afterward, all values taken are arranged in descending order, and the relationship among each value is then compared. If a value is found to be twice larger than any other value, this value is designated as be-ing long (L), and the smaller values are designated as bebe-ing short (S ). Once all relationships have been established, the average of the nine references values can be calculated and assigned to be the initial SB.
As presented in the previous example, the first nine silent values are thus 496, 470, 1102, 425, 408, 364, 386, 1087, and 1145. After sorting, the descending order of these nine silent values is 1145, 1102, 1087, 496, 470, 425, 408, 386, and 364. The first three silent values (1145, 1102, and 1087) are more than twice as large as any other value in the sample and thus are designated as L and the rest are
desig-nated as S . Afterward, the sum of long silent values (L) is divided by 3, and the sum of short silent values (S ) is calcu-lated. SB is the average value of the sum of long and short silent values. This process is illustrated by the following equation:
SB= (L/3 + S)/number of elements
SB= [(1145 + 1102 + 1087)/3 + 496 + 470 + 425
+408 + 386 + 364]/9 = 406.70
Once the initial SB value has been determined, it can be sent into the adaptive processing stage as the initial value. Meanwhile, the character detection equation can be used to calculate a subsequent SB value based on this obtained SB value to recognize spaces within elements. After a space element has been recognized, the SB value can be recalcu-lated. If the result shows L, the space element is divided by 3, and the obtained value is only then sent into the adaptive processing; otherwise, the space element is directly sent into the adaptive processing stage to obtain a new SB. Whenever a SB is obtained, the data stream is separated into elements and spaces. After the Morse code elements of a character have been isolated from a data stream, the elements can be recognized in the character recognition stage [18].
2.3 Character Recognition
Once a character space value has arrived in the tone buffer, it is a signal for the tone buffer elements to be sent to character recognition. If the recognized character set can be directly matched to a code set from the Morse code table, then it is immediately translated from the Morse code table. Oth-erwise, it has to be translated by the following minimum distance calculation. First, each tone element value in an unknown tone element stream is divided by the tone base of the previous tone element set. Then, the distances between each tone value and the code elements in each character of the Morse code table are calculated. The character with the minimum Euclidean distance to the tone value is cho-sen as the value for the unknown character. The procedure for the shortest Euclidean distance method is the following. First, each tone element, ej(xi), is divided by the tone base,
ej(xi)/tone base, for j = 1 ∼ n. Then, the roots of the sum of
the square distances between the new tone element, ej(xi),
and the character in the Morse code table are calculated. The character in the Morse code table that has the shortest Euclidean distance is recognized as the unknown character [18]. Assume, for example, the tone element values in an unknown character are the following: 878, 267, 838, and 310, in which the tone base is 287. After division by the tone base, the four tone elements are 3.06, 0.93, 2.92, and 1.08. The character with the shortest Euclidean distance in Morse code table is character “C” (i.e., 3, 1, 3, 1). Thus, character “C” is selected as the Morse code match for the unknown character.
2.4 Adaptive Processing
The variable degree variable step least-means-square (VD-VSLMS) algorithm used here serves to change the standard ‘space’ length [17]. The average of space bj(xi) (i = 1,
n− 1) in xiis the ith input data of the algorithm. The
VD-VSLMS algorithm utilizes the current data to compute a new weight vector using the weight update recursion of the stan-dard LMS algorithm with step size µ [20]. The new weight vector together with the current data are then utilized to up-date again the desired weight vector using the standard LMS algorithm weight update recursion with step size µ. Each adaptive weight vector, W(n), is adjusted according to the equation
W (n+ 1) = W (n) − α2(n) ˆ∇ (n) (4)
where
α2(n)= 2µ(1 − µXT(n)X(n)) (5)
The subscript on the α(n) is used to indicate the degree, and ˆ
∇ (n) = −2ε (n) X (n) (6)
is an estimate of the gradient. ε (n) = d (n) − XT(n) W (n)
(7) where d(n) is the scalar desired signal. µ is the step-size parameter that controls the speed of convergence as well as the steady-state and/or tracking behavior of the adaptive fil-ter. The step size µ has a value of 0.02 in our system [18]. 2.5 Support Vector Machines
Support Vector Machines (SVMs), were originally intro-duced by Vapnik and co-workers [21], [22] for classification tasks, and were subsequently extended to regression prob-lems [23]. The idea behind SVMs is the following: in-put points are mapped to a high dimensional feature space, where a separating hyperplane can be found. The algo-rithm is chosen in such a way to maximize the distance from the closest patterns, a quantity which is called the margin. SVMs are learning systems designed to automat-ically trade-off accuracy and complexity by minimizing an upper bound on the generalization error provided by Vapnik-Chervonenkis (VC) theory [25]. In this study, we applied SVMs algorithms to dots or dashes of Morse code recog-nition, which has been proven to be a powerful solution in many classification problems. However, training these sys-tems is non-trivial and a high cost of computation is required by the use of optimization packages. Kernel-Adatron (KA) algorithms [19], are used to emulate SVM training proce-dures, which are a based on the Adatron algorithm [24] but adapted by the introduction of kernels so that they can find nonlinear decision boundaries using the high-dimensional feature space, a fast and simple learning procedure which finds a maximal margin hyperplane in a high feature space.
Experimental results have shown that the predictive power is equivalent to that of a SVM and the running time can be orders of magnitude faster. The procedure of KA is shown below:
1) Initialize α0
i = 0.
2) For i= 1, . . . , m execute step 3, 4 below. 3) For a labeled point (xi, yi) calculate:
zi= m j=1 αjyjK(xi, xj) (8) 4) Calculate δαt i= η(1 − ziyi): 4.1) If (αt i+ δα t i)≤ 0 then α t i= 0. 4.2) If (αt i+ δα t i) > 0 then α t i= (α t i+ δα t i).
5) If a maximum number of iterations is exceeded or the margin λ is approximately 1 then stop, otherwise return to step 2. λ = 1 2 min {i|yi=+1} (zi)− max {i|yi=−1} (zi) (9)
3. Experimental Results and Discussion
Many individuals with motor and/or sensory disabilities are using newly-developed adapted-access software programs, hardware peripherals, and learning methods to help them use microprocessor devices via Morse code input systems through switches external to the computer. People with limited movement or sensory capabilities have been shown to successfully operate computers and other devices via adapted switching mechanisms and Morse code emulation of keyboard input functions. Research and clinical experi-ence are indicating that the fast rate of entry and low level of physical exertion inherent in a Morse code input system could make it a viable and competitive method of micropro-cessor control for persons with disabilities [12].
In this paper, an 8051 single chip is adopted to handle communication between the press-button processing and the personal computer. This recognition system is implemented using Microsoft Visual C++ 6.0, Windows version. To in-vestigate the efficiency of the proposed method, two groups of testing data, EXP1 and EXP2, were taken. EXP1 test-ing data, numbered from Exp101 to Exp115, were collected from 15 experts in the military wireless service by typing 100 identical characters. EXP2 testing data, numbered from Exp201 to Exp215, were collected from 15 abled test par-ticipants (non-experts), who were trained for a period about two months by typing 100 identical characters. The exper-imental results are shown in Table 1. The average number of matches for the EXP1 and EXP2 are 92.67 and 91.07, re-spectively. As can be expected, the experts showed a higher number of matches than the non-experts [26].
The efficiency of the proposed method was tested on three test participants with physical disabilities. Three dif-ferent test problems were used to evaluate the efficiency. Each participant typed 100 characters in 15 test samples (Dis01 to Dis15 in Table 2). The first test participant (P1)
Table 1 The recognition result for two types of test problems.
was a 14-year-old boy, who has been diagnosed with cere-bral palsy. His voluntary movements were accessible but an initial delay was exhibited before the movement was initiated, and involuntary movement partially disrupted the willed movement, making it uncoordinated. Participant 2 (P2) was a 14-year-old female adolescent, diagnosed with cerebral palsy. She exhibited involuntary and uncontrollable movements of her four limbs. The third participant 3 (P3) was a 40-year-old male adult who suffered a spinal cord in-jury with incomplete quadriparesis [18].
Experiments were conducted on a test problem set (three participants, 15 test sample for each participant). Ta-ble 2 shows that the method proposed in this study had the highest number of matches in most cases. The average num-ber of matches for SL, LVQ, and the proposed method were 73.5, 79.1, and 82.9. The reason why the proposed method elucidated a higher recognition rate can be explained in the following manner: The tone ratio (dot to dash) must be 1:3 based on the definition of Morse code. The same ratio ap-plies to dot-dash space to character space. A Morse code time series is generally an unstable one, unstable in speed and/or in rate. Maintaining precise intervals is a difficult task for persons without physical disabilities. An unstable typing speed or rate might generate two kinds of errors: tone recognition errors and space recognition errors. Therefore, it is very important to set criteria to distinguish “dot or dash” and “dot-dash space or character space.” A person’s typing rate is constant over only a short period. That means that a person’s present typing rate is similar to the typing rate of the immediately preceding several words. Because each person has his or her individual typing behavior, the dot and dash values cannot be set to a default value. In this paper, the initial criterion used to discriminate between “dot or dash” and “dot-dash space or character space” is calculated from the initial typing speed of the user. The calculation time of the current system is very short, in the range of a few mil-liseconds [9].
el-Table 2 Number of matches (out of 100) by the test participants with disabilities on test problems.
ement, the system needs to recalculate the threshold of the tone and space element. For the space recognition, the adap-tive VDVSLMS algorithm, used here improves the conver-gence characteristic of the LMS by means of gaining more accurate information about the minimum of the mean square error surface from the current data at each iteration [17]. For the tone recognition, KA is used to emulate the SVM train-ing procedure, which works by mapptrain-ing traintrain-ing data for classification tasks into a high dimensional feature space. In the feature space a maximal margin hyperplane can be found, which separates the data. After an initial dot-dash classifier is obtained from the initial training data, the test-ing data is sent to the traintest-ing process in order to obtain an adjusted decision function, which in turn increases the sys-tem’s recognition rate.
The three recognition methods were compared by ap-plying the number of matches (Table 2), in two statistical tests, the Friedman test and the multiple comparison ap-proach [27]. The Friedman test was used to test whether the total numbers of matches of the different recognition meth-ods were equal. The multiple comparison approach was used to determine which method had significantly different median total matches if the Friedman test was rejected. 3.1 Friedman Test
The Friedman test is a nonparametric counterpart of the parametric two-way analysis of variance test and was used to compare the total matches of the recognition methods when the distribution of the underlying population was not specified. The hypothesis being tested was that all the meth-ods had equal median total matches, and the alternative hy-pothesis was that all methods did not have equal median to-tal matches [27].
Let Ri j be the rank (from 1 to k) assigned to method j
on problem i. It will equal 1 if it is the lowest value among the methods. In the case of ties, average ranks are used. The test statistic can be represented by the following equation:
Tf = (n− 1) Bf − nk(k+ 1)2 4 Af − Bf (10) where Rj= n i=1 Ri j for j= 1, 2, . . . , k (11) Af = n i=1 k j=1 R2i j (12) Bf = 1 n k j=1 R2j (13)
The null hypothesis is rejected at the α significance level if the value of the test statistic exceeds the 1− α quan-tile of the F-distribution with k− 1 and (n − 1)(k − 1) degrees of freedom. We illustrate the Friedman test on the data in Table 2 (n= 15 and k = 3). The calculated value of Tf =
20.42 is greater than the critical value of F.05(2, 28)= 3.34.
We rejected the null hypothesis that all the methods had the same median total matches at a significance level α= 0.05. 3.2 Multiple Comparison Approach
The multiple comparison approach was used to deter-mine which method had significantly different median to-tal matches. Methods i and j are considered different if the following inequality is satisfied:
Rj− Ri >t(α/2)
2n(Af − Bf)/(n− 1)(k − 1)
where Ri, Rj, Af, and Bf are given previously, and t(α/2)
is a critical value on the t-table using (n− 1)(k − 1) degrees of freedom (α/2 = P(t > t(α/2))). The total matches of the three methods were ordered in an array, and a rank was as-signed to each corresponding value according to its order. The rank sums of SVM, LVQ, and SL were 42.5, 27.5, and 20.0, respectively; if the rank sums of any two methods are greater than 2.02 units apart (with α= .05), they might be re-garded as having unequal median total matches. The above numbers lead to the conclusion that the proposed method was statistically superior to both the LVQ method and SL method for the participant tested.
4. Conclusions
In this paper, a SVMs machine-learning technique was suc-cessfully applied to the problem adaptive Morse code recog-nition. Morse code has been shown to be a valuable adap-tive tool for persons with various physical disabilities and other afflictions that worsen with time and can cause a user’s ability to write, type, and speak to be progressively lost. A stable typing rate is strictly required for an accurate recog-nition of Morse code. However, maintaining precise time intervals in a Morse code time series is a challenge for a user with physical disabilities. The proposed method, which combines the support vector machines method and the vari-able degree varivari-able step LMS algorithm, was successfully applied to the problem. Statistical analyses demonstrated that the recognition rate of the proposed method was supe-rior to alternative methods discussed in the literature.
Acknowledgements
This work was supported by the National Science Council, R.O.C., under contract NSC 92-2614-E-151-001.
References
[1] R. Bower, et al., eds., The Trace Resourcebook-Assistive Technol-ogy for Communication, Control, and Computer Access, Trace Re-search & Development Center, Universities of Wisconsin-Madison, Waisman Center, 1998.
[2] D.K. Anson, Alternative Computer Access: A Guide to Selection, F.A. Davis, Philadelphia, 1997.
[3] C.-H. Luo and C.-H. Shih, “Adaptive Morse-coded single-switch communication system for the disabled,” Int. J. of Biomed. Com-put., vol.41, pp.99–106, 1996.
[4] C.-H. Shih and C.-H. Luo, “A Morse-coded recognition system with LMS and matching algorithms for persons with disabilities,” Int. J. of Medical Informatics, vol.44, pp.193–202, 1997.
[5] S.P. Levine, J.R.D. Gauger, L.D. Bwera, and K.J. Khan, “A compari-son of mouth stick and Morse code text inputs,” AAC Augmentative and Alternative Communication, vol.2, no.51, pp.45–51, 1996. [6] L.N. Goble and H.A. Colle, “High-speed Morse code training,”
Proc. IEEE 1985 National Aerospace and Electronics Conference, NAECON, pp.944–951, 1985.
[7] D.A. Shannon, W.S. Satewen, J. Miller, and B.S. Cohen, “Morse-code controlled computer aid for the non-vocal quadriplegic,” Med-ical Instrumentation, vol.15, pp.341–343, 1981.
[8] A. Thomas, “Communication devices for the non-vocal disabled,” Computer, vol.14, pp.25–30, 1981.
[9] C.-H. Yang, “An interactive Morse code emulation management sys-tem,” Computer & Mathematics with Applications, vol.46, pp.479– 492, 2003.
[10] C.-H. Yang, L.-Y. Chuang, and C.-H. Luo, “Internet access for dis-abled persons using Morse code,” International Journal of Comput-ers and Applications, vol.26, no.1, pp.10–16, 2004.
[11] C.-H. Yang, L.-Y. Chuang, and C.-H. Luo, “Morse code application for wireless environmental control system for severely disabled in-dividuals,” IEEE Trans. Neural Systems & Rehabilitation Engineer-ing, vol.11, no.4, pp.463–469, Dec. 2003.
[12] T.W. King, Modern Morse Code in Rehabilitation and Education, Allyn and Bacon, MA, 1999.
[13] T. Friess, N. Cristianini, and C. Campell, “The kernel-adatron algo-rithm: A fast and simple learning procedure for support vector ma-chines,” Proc. fifteenth International Conference on Machine Learn-ing, ICML, pp.180–196, 1998.
[14] D. Roobaert, “Improving the generalization of support vector ma-chines: An application to 3D object recognition with cluttered back-ground,” Proc. Workshop on SV Machines at IJCAI, pp.29–34, 1999.
[15] R. Rychetsky, S. Ortmann, and M. Glesner, “Support vector ap-proaches for enigner knock detection,” International Joint Confer-ence on Neural Networks, 1999.
[16] B. Scholkopf, Support vector Learning, Ph.D. Dissertation, Max-Planck-Institut fur biologische Kybernetik, 1997.
[17] M.A. Khasawneh and K.A. Mayyas, “A newly derived variable de-gree variable step size LMS algorithm,” Int. J. Electronics, vol.79, no.3, pp.255–264, 1995.
[18] C.-H. Yang, “Morse code recognition using learning vector quanti-zation for persons with physical disabilities,” IEICE Trans. Funda-mentals, vol.E84-A, no.1, pp.356–362, Jan. 2001.
[19] T. Frieß, N. Cristianini, and C. Campbell, “The Kernel-Adatron: A fast and simple learning procedure for support vector machines,” Proc. Fifteenth International Conference on Machine Learning, pp.188–196, 1998.
[20] B. Widrow and S.D. Stearns, Adaptive Signal Processing, Prentice Hall, Englewood Cliffs, NJ, 1985.
[21] B. Boser, I. Guyon, and V. Vapnik, “A training algorithm for opti-mal margin classifiers,” Fifth Annual Workshop on Computational Learning Theory, ACM Press, 1992.
[22] C. Cortes and V. Vapnik, “Support vector networks,” Mach. Learn., vol.20, pp.273–297, 1995.
[23] H. Drucker, C. Burges, L. Kaufman, A. Smola, and V. Vapnik, “Sup-port vector regression machines,” in Neural Information Processing Systems, eds. M. Mozer, M. Jordan, and T. Petsche, vol.9, pp.155– 161, MIT Press, Cambridge, MA, 1997.
[24] J.K. Anlauf and M. Biehl, “The Adatron-an adaptive perceptron al-gorithm,” Europhysics Letters, vol.10, pp.687–692, 1989. [25] V.N. Vapnik, The Nature of Statistical Learning Theory, Springer
Verlag, 1995.
[26] C.-H. Yang, S.-H. Shieh, and L.-C. Jin, “Morse code recognition using support vector machines,” Proc. 2003 IEEE-EMBS Asian-Pacific Conference on Biomedical Engineering, Keihanna, Japan, Oct. 2003.
[27] W.J. Conver, Practical nonparametric statistics, 2nd ed., Wiley & Sons, New York, 1986.
Cheng-Hong Yang is a full professor of the Electronic Engineering Department of National Kaohsiung University of Applied Sciences. He obtained M.S. degree and Ph.D. degree in com-puter engineering from North Dakota State Uni-versity, in 1988 and 1992, respectively. His main areas of research are evolutionary compu-tation, bioinformatics, and assistive tool imple-mentation.
Li-Yeh Chuang received her M.S. de-gree from the University of North Carolina at Greensboro, in 1988, and Ph.D. degree from North Dakota State University, in 1994. She is an associate professor in the department of chemical engineering, I-Shou University. Her current research interests are rehabilitation engi-neering, biochemistry, bioinformatics, and mul-timedia systems.
Cheng-Huei Yang received the B.S. degree in 1978 and the M.S. degree in 1987 from Na-tional Taipei Institute of Technology and North-eastern University, respectively. He received his Ph.D. degree in electrical engineering from Na-tional Cheng Kung University in 2001. Cur-rently, he is an associate professor in the Depart-ment of Electronic Communication Engineer-ing, National Kaohsiung Marine University. His current research interests are network commu-nication, electronic instrument systems, and im-age processing.
Ching-Hsing Luo received the B.S. de-gree in electrophysics from the National Chaio Tung University and the M.S. degrees in electri-cal engineering from the National Taiwan Uni-versity in 1982 and in biomedical engineering from the Johns Hopkins University in 1987. He received the Ph.D. degree in biomedical engi-neering from the Case Western Reserve Univer-sity in 1991. He is a full professor in the Depart-ment of Electrical Engineering, National Cheng Kung University in Taiwan since 1996. His re-search interests include biomedical instrumentation, assistive tool imple-mentation, cell modeling, signal processing, home automata, and quality engineering.
