GB2199976A - Automatic pattern recognition - Google Patents

Abstract

In a parallel processing system, a sequence of input data e.g. a line of printed characters is scanned by a multi-channel sensor 14. Whenever a data element is detected in any channel an associated pulse generator 25-29 produces a short pulsed signal, the recurrence frequency of which uniquely identifies the channel concerned. The pulse signals are input discretely to transmission channels of an array 32, comprising similar monostable delay units 40 each providing a nodal output point for interrogation or interconnection. Transmission proceeds stepwise along the any so that a pattern of pulse signals analogous to the pattern of data is produced. A matrix of nodal points is designated for each character or pattern and the detection of simultaneous signals at all points of the matrix indicates successful recognition. Recognition signals, again in coded pulse form, are similarly combined to build up words or larger patterns indefinitely. Other applications include patterns of industrial process control and neural modeling. In an automatic welding system the multi-channel sensor (115, 117, Fig 6) senses position of a workpiece (110) and pressure between welding electrodes (113, 114). On recognition of the desired pattern power is supplied to a magnetic-clamp (116) and the welding electrodes (113, 114). <IMAGE>

Description

AUTOMATIC PATTERN RECOGNITION The invention relates to the automatic recognition of a pattern in a sequence of data. Visual and auditory data represent examples of particular interest, in the recognition of printed characters and of speech, but there is a very general range of application to situations involving an automatic response to the presentation of data which, it is known or suspected, may include recurrent or prescribed forms or patterns.

In accordance with one aspect of the invention there is provided apparatus for the automatic recognition of a pattern irt a sequence of data comprising sensing means arranged for the simultaneous monitoring of a plurality of blocks of data elements in the sequence to detect those elements within each block in predetermined order, means responsive to the detection of each such element to generate a signal comprising a sequence of pulses and having a pulse frequency which is assigned to identify the block from which the signal is derived, means providing in an array of transmission channels at least one discrete channel for each signal frequency, means for detecting the passage of signal pulses in each such channel, and recognition means responsive to the detection of appropriately related signal pulses to produce a recognition signal indicating recognition of a desired pattern.

Preferably, each transmission channel comprises a like succession of retransmission delay means having respective nodal output points at which the passage of signal pulses may be detected, a predetermined matrix of such nodal points being allocated to the recognition of a desired pattern.

The recognition means may have associated storage and control means operative to determine each pattern to be recognised and the matrix of nodal points to be allocated to the recognition of each such pattern.

Each recognition signal may comprise a pulse sequence of characteristic frequency. Further respective transmission channels may be provided to receive a plurality of such signals, thereby enabling a desired concatenation or combination of patterns to be recognised at a further matrix of nodal points.

The recognition means may comprise counting means arranged to determine the rate of passage of pulses at each nodal point of the matrix. Preferably the recognition means includes means for comparing the pulse rate at each nodal point of the matrix with an appropriate reference pulse signal.

A retransmission delay means may include disabling means responsive to-a signal from recognition means associated with a preceding nodal point to inhibit retransmission of pulse signals representing a pattern which has been recognised.

In an apparatus for character recognition, or other application in which the sequence of data is pictorial, the sensing means may include line-scan means and each block of data may comprise the contents of a single line so scanned.

In such apparatus the sensing means is arranged to scan a plurality of lines simultaneously and the corresponding inputs to the array of transmission channels may be arranged to preserve in the resulting pattern of pulse signals the spatial relationship between elements of successive lines.

In an apparatus for speech recognition, or other application in which the sequence of data represents sound, the sensing means may include filter means for isolating selected frequency bands in the data and each such band may constitute a block of data. The corresponding inputs to the array of transmission channels may then be arranged to preserve in the resulting pattern of pulse signals the temporal relationship between elements of successive bands.

In an apparatus arranged for the control of an industrial process, the sensing means may be operative to sense the value of each of a plurality of parameters of the process and the pattern to be recognised may be a combination of such values which establishes the condition for a further step in the process, the recognition signal being arranged to enable such step.

In such apparatus, the array of generally forwardly directed transmission channels may include at least one pair of channels arranged for transmission in the reverse direction, each such channel being coupled to a forward channel to provide a reflected signal path, and the recognition means being arranged to compare at least said pairs of forward and reflected signals to establish a condition for enabling output from the reverse channels.

In accordance with another aspect of the invention, a method for the automatic recognition of a pattern in a sequence of data comprises the operations of sensing simultaneously a plurality of blocks of data elements in the sequence to detect in predetermined order those elements within each block, generating in response to the detection of each such element a signal comprising a sequence of pulses and having a pulse frequency which is assigned to identify the block from which the signal is derived, providing in an array of transmission channels at least one discrete channel for each signal frequency, each channel being so arranged that transmission proceeds stepwise along a like succession of nodal points whereby a pattern of pulse signals analogous to the pattern of data is produced, and monitoring a selected matrix of nodal points to detect the occurrence of a desired pattern.

It is relevant to an appreciation of the invention to refer to some current views on the organisation of the brain. It is commonly understood that although the brain has a great facility for the rapid recognition of patterns, the processing of incoming data depends on the response of nerve cells to sensory signals, called action potential or spikes, which appear to recur at a quite low frequency in the order of lkHz. This limitation in speed is said to be overcome by the availability of very large numbers of cells for parallel processing of the data.

The present inventor further emphasises that not only are many items of data processed in parallel but each item may be widely distributed to be tested simultaneously for a number of different attributes, for example, or for comparison with memory. In considering the possible manner and means of such data distribution, he has appreciated the value of a general computing facility providing parallel processing of input data which is coded in terms of the frequency (pulse rate) of a repetitive pulse signal. Such an apparatus offers a valid analogue for the study of cerebral behaviour and becomes capable of a range of tasks which may be generally classified as automatic pattern recognition.The term 'pattern' is to be understood very broadly and extends from symbols (alphabetic, chemical or mathematical for example), pictures and patterns of sound to the modelling of suitable practical or abstract problem situations which may then be resolved.

Such scope of application derives from a procedure which is consistent with the nature of the 'spike' or action potential, which represents a complex probability, and also with Shannon's theory of communication. The theory deals with the serial transmission of information in adverse conditions such that the outcome involves the summing of a series of probabilities (expressed as logarithms to base two) to assess the validity of a number of items of data. It will be seen that the present approach depends on the substitution of a log(frequency) term for log(probability) and that the summation is carried out in parallel. The frequency term in fact defines an average frequency of occurrence, equivalent to a value of probability confirmed by experience. The use of such frequency labelling of signals not only provides substantial immunity to noise but also enables the signal to be related to its origin, however far the signal transmission network may be extended.

In a preferred form of the invention, the selection for interrogation of a particular matrix of nodal points in the array of transmission channels determines the pattern to be detected. In the case of character recognition, the selection of points to form each matrix will be decided from a knowledge of the style of character to be encountered or an approximation to it. The brain of an animal, however, will depend on stored matrices which correspond to various sensory stimuli, an immediate response to which may govern success or survival. In a similar way, the inventive concept includes the option of a storage and control facility so that new matrix forms may be selected as required. Thus selection might proceed in response to an external signal indicative of the general nature of the data to be presented.It is also possible to arrange that a previously unrecognised pattern, that is, one not in the current matrix repertoire, should be entered in the memory. This might be done in response to a stimulus which, for example, may naturally occur simultaneously with the appearance of the pattern. This action is a simple form of adaptive, or learning, response.

Alternatively, if the previously unrecognised pattern is repeated a significant number of times within the timescale of observation, it may be detected as an unexpected regularity of signal distribution within a block of nodal points which is monitored for the purpose. A form of histogram is built up which enables even a rarely occurring pattern to be detected on a sufficiently extended timescale of observation.

An embodiment of the apparatus of the invention, and its method of use, will now be described by way of example with reference to the accompanying drawings in which : Figure 1 represents schematically the input stage of a character recognition apparatus in accordance with the invention; Figure 2 represents schematically a recognition circuit for use with the apparatus of Figure 1; Figure 3 represents schematically a frequency comparison arrangement for the circuit of Figure 2; Figures 4a and 4b relate to the operation of the circuit of Figure 3; Figure 5 represents schematically a further stage in the apparatus of Figure 1; and Figure 6 represents schematically a modified form of the apparatus in use for the control of an industrial operation.

With reference to Figure 1, a paper sheet 10 bearing printed characters, such as a character 12 (letter E, shown hatched), is arranged for scanning from left to right by a multichannel optical sensing heaS 14. Sheet 10 is so orientated, if necessary with the aid of a preliminary scanning device (not shown), that the axis containing the sensors of head 14 is parallel to the vertical edge of the letter E, the scan axis being parallel to the horizontal edge of letter E. The resolution of the sensors of head 14 and the size of the letter are such, in this example, that the letter may be considered as comprised of five elements in height. All five elements are sensed simultaneously with each-element corresponding to a discrete channel of head 14. In descending order, the elements of each scanning line are indicated as assigned to sensors 15 to 19 respectively.The scan proceeds in steps horizontally such that the full width of the letter is observed as four discrete sequential elements during equal intervals of time. In general, the elements observed during the scan of any single line will be said to comprise a block of data.

A response signal is produced in the appropriate channel of head 14 whenever an inked element of letter E is exposed to a sensor. Each sensor 15 to 19 communicates with a similarly ordered one of a set of pulse generators 25 to 29. The function of each generator is to produce, in response to an input from the related sensor, a signal comprising a short sequence of pulses having a repetition rate proportional to the logarithm of a distinctive predetermined frequency. Thus a frequency band or a mid-point frequency is uniquely assigned to each block of data.By analogy with a pattern of sound in the audio range (extended for arithmetical simplicity), the elements of the top scan line of character 12 are to be represented by a frequency of 100kHz and the succeeding lines by lOkHz, lkHz, 100Hz and 10Hz respectively. Taking logarithms, and again for simplicity of explanation base ten is used rather than base two, the signal rates required from generators 25 to 29 become respectively 5, 4, 3, 2 and 1 pulses per unit of time.

The signal from each generator is input to a respective one of an array 32 of transmission channels. As indicated in Figure 1, generators 25 to 29 feed channels 35 to 39.

Each channel of array- 32 comprises a succession of non ivertina monostable delay units such as unit 40 which have similar nominal delay characteristics. Pulses on different channels are thus retransmitted at each delay unit to advance at substantially similar stepping rates. The time relationship between channels is therefore preserved, within the limits of variation between different delay units. Diagrammatically, the channels being shown as parallel, the picture elements from different scan lines represented by their respective pulse signals will maintain their correct positional relationship. The transmission of a pulse signal is detected at the output point of any delay unit, which, since it provides for interconnection, will be referred to as a nodal point or node.Thus, in accordance with each character to be detected a corresponding selection or matrix of nodes is designated for external connection to a recognition circuit (shown in Figures 2 and 3). For the letter E, the matrix comprises nodes 41 to 44 in channel 35, node 45 in channel 36, nodes 46,47 in channel 37, node 48 in channel 38, and nodes 49 to 52 in channel 39. It is convenient, but not essential, for the matrix to be (as shown here) a reversed representation of the character since in the sequence of pulses along the transmission channels those derived from detection of the left-hand elements of the character will precede those from the right. To consider the point further: although the function of the matrix is more easily understood in pictorial terms, and will be explained in that way, it is not necessary for the outline of the matrix to be modelled on that of the character. Any nodes on appropriate channels can be used provided the correcttime relationship is observed between the various pulse trains.

In operation, if the pattern of pulses derived from letter E appears on array 32 it will be detected at the E matrix. The pulses then continue to step along their respective channels and may or may not be useful for further reference. A pulse pattern derived from a different character will produce no response at these matrix and will continue to step along the array until it reaches a matching matrix.

Referring now to Figure 2, a portion of array 32 is shown which includes the recognition matrix for letter E anodes 41 to 52. In the design of a suitable detection circuit it is necessary to consider the transmission time of pulses through the matrix. Suppose the delay at each node to be of duration d. Then the interval between pulses must be greater than d; say 2d. Channel 35 carries five pulses for each element of the upper bar of letter E and so determines that an interval of 10d must be allowed at each node for each element or 40d for the transmission of the complete character. Recognition therefore requires observation for a period of 40d during which pulse trains should reach the correct nodes at the correct intervals.

Clearly a partial match will be given by the cutters I, F or L and all identifying information must be collected.

Thus, if pulses arrive at all the nodes 44 to 52 of the right-hand column, a decision on their significance must be delayed for 30d; for nodes 43,46,51 the delay is 20d and for nodes 42,50 it is 10d, after which the status of nodes 41,49 will be known and a positive identification is possible.

This procedure is followed in a recognition circuit in which the pulse signal at the output of each of nodes 44, 45,47,48,52 is taken individually to a unit 54 which operates to verify the nature and simultaneous presence of the signals and to produce an output signal delayed by 30d.

A similar unit 56 is arranged to receive signals from nodes 43,46,51 and to produce an output delayed by 20d; a unit 58 responds to nodes 42,50 with a delay of lOd; and a unit 60 responds to nodes 41,49 without delay. The outputs from units 54,56,58,60 are finally tested for coincidence- in a unit 62 which produces an output signal suitable for the operation of a1printer or other means for indicating the identification of a character.

With reference to Figure 3, unit 54 will be described in more detail. In order to test that the pulse rate at node 44 is correct, the node signal is input to an AND gate 64 together with a reference signal REF(44) having the required pulse rate of 5 pulses/unit time. The output signal is divided by five in a diver 65 to give a single pulse for input to an AND gate 66. The signals from nodes 45,47,48,52 are processed similarly, the corresponding reference signals being of appropriate pulse rate and the divisors being of appropriate value. In summary, REF(45), REF(47), REF(48), and REF(52) have respective pulse rates of 4,3,2 and 1/unit time for input to AND gates 67,68,69 and 70.The related dividers 71,72,73 have divisors of 4, 3 and 2, AND gate 70 having a single pulse output and requiring no divider. The output from AND gate 66 signifying the coincidence of pulse signals at the five nodes is delayed by a period 30d at a unit 74 and the output from unit 74 is input t6 the final AND gate 62 as described with reference to Figure 2.

Further details of the arrangement for frequency comparison, indicated by the REF IN inputs of Figure 3, are now described with reference to Figures 4a and 4b.

Consideriny the input from node 44 to AND gate 64, as ir, Figure 3, the simplest form of comparative input to REF IN is a pulse train of identical frequency. Figure 4a shows an improved form in which gate 64 has either two or three inputs in addition to that from node 44. On lines 76,77 reference signals are applied which are similar in frequency but respectively slightly retarded and slightly advanced in phase with respect to the signal from node 44.

In Figure 4b, waveform A is the input from node 44, waveform B is the input on line 76, waveform C is the input on. line 77 and waveform E is the output from gate 64.

High selectivity is shown in passing the desired frequency. Waveform E also demonstrates a modified output achieved by applying on a line 78 a further input waveform D which comprises sequences of pulses at the desired frequency with an interval between each sequence and the next. The effect is to disable gate 64 for controlled periods.

The selection of each frequency of reference waveform to be generated is controlled from stored data relating to a particular matrix. Synchronism is achieved by directly matching the rate of advance of reference pulses to that of signal pulses from transmission array 32. For this purpose reference pulses are derived from a recirculating loop containing one or more delay monostables similar to those at the nodes of array 32. The shortest circulation time, corresponding to the highest frequency, occurs in a loop containing only one such node. The frequency is stepped don by inserting further nodes until the lowest frequency is obtained with five nodes in the loop. When the reference pulse rate and the signal pulse rate are exactly matched, the arrangement is in effect a pulse counter.As an alternative to phase variation, selectivity is similarly improved by slightly varying the reference frequencies on lines 76,77 above and below the signal frequency. This i achieved by slight modification of the delay characteristic of the monostables in the respective recirculating loops.

With further reference to Figure 2, it was shown that the the output from gate 62 may be used as the final output to a printer, for example. In a further development of the apparatus, the output signal from gate 62 is shown in Figure 5 to be input to a pulse generator 80 (similar to the pulse generators 35 to 39) which produces a signal at a pulse rate which is assigned to represent uniquely the complete letter E. Similarly, the final outputs at gates 81 to 85 from recognition matrices for other complete characters are converted to respective identifying pulse frequencies at generators 82 to 86.

The pulse signal representing E is input to a transmission channel 90 which comprises a sequence of monostable delay nodes similar to the channel 35 of Figures 1 and 2. Other signals representing, say, the letters A, T and N are input to similar channels 92 to 96 forming an array 98.

Characters which were read in succession by sensing head 14 will be signalled in similar succession along the channels of array-98. Thus, as a result of a continuous scan of the nodes of array 98 or by the use of a recognition matrix at nodes 100 to 104, the complete word EATEN can be detected. Such a process of concatenation may be extended as required to assemble words or images.

Referring again to the description relating to Figure 1, it has been assumed that pulse signals representing the elements of letter E will continue to step along the channels of array 32 after recognition at the E matrix.

In some cases it will be preferred to inhibit retransmission of such pulses. The output signal from gate 62, or a further signal derived from it, is then applied to disable the next succeeding delay units for a period equal to 40d, during which all E pulses will be lost.

The invention is considered to provide an advantageously simple approach to the design and use of pattern recognition apparatus in that information is generated and transmitted in the encoded form of pulse repetition frequency, which is always readily identifiable.

Processing depends largely on the simple AND function and a systems therefore capable of expansion to a large capacity. Flexibility is retained since a memory holding selectable values of frequency is sufficient to change the operation of a recognition matrix. A form of apparatus has been described in terms of discrete device circuits but, dependent on the scale and complexity of any proposed apparatus employing electrical signalling, other arrangements including VLSI may be suitable. The future use of optical signaling with appropriate forms of transmission and recognition apparatus is also visualised.

The preceding description with reference to the drawings has related to character recognition and in this connection -it should be noted that there is no restriction on the form of the character, since the form to be identified is only determined in setting up the recognition matrix. It is also intended that the recognition of patterns in any suitable sequence of data, including sound patterns such as speech, should be within the scope of the invention. The input sensing stage must of course be appropriately adapted. For a sound pattern, filters would be used to divide the input into a series of frequency bands, each comprising a block of data and each to be represented by a unique value of pulse repetition rate. The manner of processing the pulse signals, as explained for characters, ensures that temporal relationships are preserved.In the basic case, no account is taken of variations in amplitude but such an additional dimension in the data is accommodated by correspondingly coded additional channels in the recognition array.

Similarly, for characters or pictorial images, variations in brightness or colour could be encoded.

It is expected that, since the design is based on a view of neurological mechanisms, the apparatus will be useful in modelling cerebral behaviour. For example, the principle appears to be applied in nature that the most common, or most urgent, sensory input is to be processed at the earliest possible stage in the neural network.

Thus, the initial processing of visual signals is carried out at the retina and the identification of scent by insects may be localised in the antennae. The choice of the letter E in explaining the working of the apparatus is a simple analogous application of the natural principle, since E is the most frequently occurring letter in European texts and is recognised at the earliest stage of the array 32. The scope for simulation of biological systems can be seen to be extensive in that natural evolution depends on the recognition and retention of patterns of information from the environment which are critical to survival.

In a similar manner, a machine incorporating memory may be made to respond to its environment and, as a result of experience, to select only those responses which lead to approval. Industrial control may thus be regarded as being of the same class as life processes, characterised in organisms by recursive memory; that is, the whole organism derives in some way from all the previous states constituting its evolutionary path. The process may be visualised in an industrial control apparatus which includes an array of transmission channels, similar to the array 32 of Figure 1, each comprising a succession of nodal points from which recognition matrices are selected in various configurations. Consider a simplified form of automatic or robotic welding operation as shown in the schematic diagram of Figure 6.A workpiece 110 is carried by means of a conveyor system from left to right along an axis 112 towards a welding position between a vertically movable electrode 113 and a stationary electrode 114. The progress of the workpiece is monitored by a sensor 115 until it reaches the welding position where it is to be held by an electromagnetic clamp 116. The position of electrode 113 is similarly monitored and the optimum contact position is finally determined in terms of the pressure between electrodes 113 and 114 by means of a pressure sensor 117. The positioning sensors will generate respective blocks of numerical data of which only one element in each block will have the specific predetermined value to cause a detection signal to be generated by means of a respective pulse generator (not shown).

Figure 6 also shows a portion of a transmission array 119 in which each of a group of six channels 120 to 125 is indicated by a broken line, those signal paths of present interest being shown by full lines. Channels 122 and 125 are supplementary to the basic array in that their direction of transmission is reversed with respect to the main channels i.e. from right to left in the diagram. In operation, channel 120 receives from sensor 115 a signal in suitably pulse-frequency coded form which indicates the arrival of the workpiece at the welding position. The signal is caused to step along channel 120 by delay units 126 to 129 until at unit 130 it is diverted to unit 131 in channel 121. The signal is then reversed to travel along channel 122 via delay units 132 to 136 for eventual output to magnetic clamp 116. In a similar manner, a second input signal, which indicates that the electrode contact pressure is correct, is received on channel 123 from sensor 117. This signal is stepped along channel 123 by delay units 137 to 141, reversed at unit 142 on channel 124 and returned via units 143 to 147 along channel 125 for output to a control mechanism 148 for the power supply to the welding electrodes.

The output signals on channels 122 and 125 are now subject to the operation of respective switches 149 and 150, which in turn depend on an inspection of the pattern of signals by reference to an acceptance condition held in the machine memory. The inspection is carried out at nodal points 128, 134, 139 and 145 by connecting these points to an AND gate 151. In practice, delay and division stages are necessary, similar to those of Figure 3, but are omitted for simplicity of description. Such stages are required to accommodate the transmission times round the loops 128, 131, 134 and 139, 142, 145 and the difference in pulse count between the signals in the two loops.

Taking these factors into account, when the input and output loop signals are all present at the four nodal points an output is produced at gate 151 which is caused to actuate switches 149, 150. The output signals on channels 122 and 125 then enable the welding operation by causing the workpiece to be held in place while the electrodes are energised.

A simplified operation involving the coordination of only two sensory functions and two motor functions at a single location has been explained in detail in order to establish clearly a procedure which is capable of indefinite extension. For example, signals from other sources may be input to the array to be modified or modulated by inputs from the sensors and the operation of actuators can be initiated or modified by such signals.

In a representation of a complex situation, the number of sets of signal loops and the corresponding number of AND gates similar to gate 151 can be made large and their outputs can in turn be combined in the manner of the image concatenation procedure described with reference to Figure 5. The number of signal loops within each set may also be increased. Such extension enables the modelling of an industrial operation having a network of inter-related functions or of the action of the brain in coordinating multiple sensory inputs with one or more motor responses.

It is also envisaged that selected matrices of nodal points may be so configured and modulated that they could be incorporated in the design of computational cellular automata. The term 'cellular automaton' relates to a twoor three-dimensional array of cells having the property that data may be propagated from any cell, via intermediate cells, to more remote cells in the system. A switching capability can be used to establishidynamic, recirculating memory and the device has some of the properties of a computer. Such automata are expected to be capable of application in the following fields a. Molecular biology, particularly the mechanism of action of the genetic code in relation to embryology, synthesis of proteins and co-operativity of enzymes.

b. General biology, particularly in the field of evolutionary studies.

c. Theoretical physics, particularly certain aspects of quantum mechanics.

d. Industrial control processes, including robotics, and in relation to the improvement of safety features and the use of biosensors.

e. Medicine, particularly in modelling the overall chemical and electrical activity of the central nervous system and in relation to the mechanisms of action of natural and synthetic therapeutic substances.

f. Environmental studies, particilarly in modelling large and complex systems.

g. Econometric studies, particularly in modelling large and complex systems.

h. Psychology and sociology, particularly in analysing the behaviour of the individual in relation to that of groups and their environments.

j. Where it is relevant to an application, groups of selected matrices of nodal points may be configured to represent the reflecting, refracting and focusing properties of materials in the visible and other regions of the electromagnetic spectrum.

A further point of relevance to such applications arise from Figure 6 and the accompanying discussion concerning the operation of the signal loops, suchas the the loop 126 to 130, 131, 132 to 136. The reversal of the signal path at the central point 131 may be considered as a reflection which provides an opportunity for selfcomparison i.e. a comparison between the original and current states of the signal. Such self-comparison is an aspect of the recursive memory (to which reference was made earlier in this specification) and as such is considered to be an essential feature of the successful evolution, past and present, of biological and physical systems. This view is consistent with modern developments in phase conjugation in holography and in the treatment of the mathematical class of NP-complete'problems (such as the 'travelling salesman' problem) requiring explicit simulation of all possibilities for their solution. For these reasons the listed fields of interest are expected to be amenable to modelling and analytical study by means of the approach to automatic pattern recognition comprised in the present invention.

Claims (17)

1. Apparatus for the automatic recognition of a pattern in a sequence of data comprising sensing means arranged for the simultaneous monitoring of a plurality of blocks of data elements in the sequence to detect those elements within each block in predetermined order, means r-esponsive to the detection of each such element to generate a signal comprising a sequence of pulses and having a pulse frequency which is assigned to identify the block from which the signal is derived, means providing in an array of transmission channels at least one discrete channel for each signal frequency, means for detecting the passage of signal pulses in each such channel, and recognition means responsive to the detection of appropriately related signal pulses to produce a recognition signal indicating recognition of a desired pattern.

2. Apparatus as claimed in Claim 1 in which each transmission channel comprises a like succession of retransmission delay means having respective nodal output points at which the passage of signal pulses may be detected, a predetermined matrix of such nodal points being allocated to the recognition of a desired pattern.

3. Apparatus as claimed in Claim 2 in which the recognition means includes storage and control means operative to determine each pattern to be recognised and the matrix of nodal points to be allocated to the recognition of each such pattern.

4. Apparatus as claimed in any preceding claim in which each recognition signal comprises a sequence of pulses having a recurrence frequency which is assigned to identify the pattern to which the signal relates.

5. Apparatus as claimed in Claim 4 comprising a further similar array of transmission channels to receive discretely a plurality of recognition signals and further recognition means enabling the recognition of a desired concatenation or combination of patterns.

6. Apparatus as claimed in any of Claims 2 to 5 in which the recognition means comprises pulse-counting means arranged to respond to the presence at each nodal point of an appropriate matrix of a signal of respectively predetermined pulse frequency.

7. Apparatus as claimed in any of Claims 2 to 5 in which the recognition means comprises means for generating at least one pulsed reference signal the pulse rate of which is related to that of the signal to be detected at a nodal point in a recognition-matrix and means for comparing the reference signal with the signal derived from that point.

8. Apparatus as claimed in any preceding claim comprising, in cooperation with retransmission delay means, disabling means responsive to a signal from recognition means operative at a preceding nodal point to inhibit retransmission of pulse signals representing a pattern which has been recognised.

9. Apparatus as claimed in any of Claims 2 to 8 in which the retransmission delay means comprises non-inverting monostable delay means.

10. Apparatus as claimed in anyApreceding claim and arranged for the recognition of printed characters or pictorial matter in which the sensing means comprises line-scan means and each block of data comprises-the contents of a single line so scanned.

11. Apparatus as claimed in Claim 10 in which, for the processing of a plurality of line-scan signals, the corresponding inputs to the array of transmission channels are arranged to preserve in the resulting pattern of pulse signals the spatial relationship between elements of successive lines.

12. Apparatus as claimed in any of Claims 1 to 9 and arranged for the recognition of patterns of sound, including speech, in which the sensing means includes filter means for isolating selected frequency bands, each such band constituting a block of data.

13. Apparatus as claimed in Claim 12 in which, for the processing of a plurality of signals derived from such bands, the corresponding inputs to the array of transmission channels are arranged to preserve in the resulting pattern of pulse signals the temporal relationship between elements of successive bands.

14. Apparatus as claimed in any of Claims 1 to 9 and arranged for the control of an industrial process in which the sensing means is operative to sense the value of each of a plurality of parameters of the process and the pattern to be recognized is a combination of such values which establishes the condition for a further step in the process, the recognition signal being arranged to enable such step.

15. Apparatus as claimed in Claim 14 in which the array of generally forwardly directed transmission channels includes at least one pair of channels arranged for transmission in the reverse direction, each such channel being coupled to a forward channel to provide a reflected signal path and the recognition means being arranged to compare at least said pairs of forward and reflected signals to establish a condition for enabling output from the reverse channels.

16. Apparatus for the automatic recognition of a pattern in a sequence of data substantially as described herein with reference to and as shown in the accompanying drawings.

17. A method for the automatic recognition of a pattern in a sequence of data comprising the operations of sensing simultaneously a plurality of blocks of data elements in the sequence to detect in predetermined order those elements within each block, generating in response to the detection of each such element a signal comprising a sequence of pulses and having a pulse frequency which is assigned to identify the block from which the signal is derived,providing in an array of transmission channels at least one discrete channel for each signal frequency, each channel being so arranged that transmission proceeds stepwise along a like succession of nodal points whereby a pattern of pulse signals analogous to the pattern of data is produced, and monitoring a selected matrix of nodal points to detect the occurrence of a desired pattern.