Pattern Recognition System

The present invention relates to generally a pattern recognition system and more particularly a pattern recognition system capable of identifying the input patterns correctly even when an input pattern has a strong similarity with more than one standard patterns in storage for identification.

In a conventional superposition type pattern recognition system, an input pattern is superposed upon each of standard patterns of different categories in storage so that there may be established one-to-one correspondence between the input pattern and a number of the standard pattern which has the maximum similarity to the input pattern. The pattern recognition system of the type described is simple in construction so that it has been widely used in various fields especially when the characters and numerals to be identified have been standardized. The superposition type pattern recognition system is also used in order to identify the characters and numerals written within a predetermined space or block, but an additional operation for normalizing the deviation of a character or numeral from the standard is required. Furthermore, in the superposition type pattern recognition system, the overall input and standard patterns are compared with each other in order to determine the degree of similarity therebetween. Therefore, when the overall input and standard patterns are similar macroscopically, the similarity of the input pattern to the standard pattern is determined as strong even when they are apparently different from each other in minute points. As a result, an input pattern is identified as having strong similarity to two or more standard patterns so that identification of the input pattern becomes impossible.

One of the objects of the present invetnion is therefore to provide an improved pattern recognition system capable of correctly identifying an input pattern even when it exhibits strong similarity to more than one standard pattern of different categories in storage.

In general a human observer identifies a pattern such as a character first by discerning the overall configuration of this pattern and then distinguishing the specific features of the pattern when he sees that its overall configuration has strong similarity to more than one pattern so as to reach the final conclusion. The present invention may simulate this humane pattern recognition process.

According to one embodiment of the present invention, an input pattern is converted by an input device into electrical video signals, which are quantized and stored in a two-dimensional memory. The input pattern information stored in the memory is converted into a plurality of analog voltages representing the features of the input pattern. The analog voltages are applied to a similarity identification unit so that the degree of similarity of the input pattern to the standard patterns in storage may be obtained. In accord with the present invention, the similarity of the input pattern to each of the standard patterns is represented by an analog voltage. The analog output voltages are applied to an identification unit so that there may be established one-to-one correspondence between the input pattern and one of the standard patterns when only one analog output voltage representing said one standard pattern is in excess of a predetermined level. The above steps are similar to those of the conventional pattern recognition systems. According to the present invention, when the input pattern exhibits strong similarity to two of the standard patterns in storage, an input pattern enlargement control unit is actuated to enlarge a part of the input pattern so that the latter may be identified as belonging to one of the two standard patterns based upon the features extracted from the enlarged input pattern. However when the input pattern exhibits similarity to three or more standard patterns in storage, a signal representing the impossibility of identification of the input pattern (or rejection thereof) is generated.

The above and other objects, features and advantages of the present invention will become more apparent from the following description of one preferred embodiment thereof taken in conjunction with the accompanying drawing.

FIG. 1 is a block diagram of one preferred embodiment of a pattern recognition system in accordance with the present invention;

FIG. 2( a) shows one example of an input pattern stored in a two-dimensional memory thereof;

FIG. 2 (b) shows an enlarged part thereof;

FIG. 3 is a circuit diagram of an input pattern identification unit; and

FIG. 4 is a fragmentary circuit diagram of an auxiliary input pattern identification unit.

Referring to FIG. 1, an input unit 2 reads a pattern 1 and converts it into electrical signals in such a way that the digital 1 signal represents a black elementary area while the digital 0 signal represents a white elementary area. The digital video signals are transferred into a two-dimensional memory 3. The input unit 2 is of a conventional type comprising, for instance, a flying spot scanner or an array of photoelectric cells and a quantizer adapted to convert the output of the former into the electrical video signals in the manner described above. The two-dimensional memory 3 comprises a plurality of shift registers arrayed two-dimensionally. Since the input unit 2 and the two-dimensional memory are well known to the art further description thereof will not be given in this specification.

The output of the memory 3 is transferred into a pattern component extracting unit 4 which is adapted to divide the input pattern information read out from the memory 3 into a number of n regions and to provide the pattern component signals each representing each of n regions which in turn are applied to a number of n linear adder circuits. Each of the latter circuits is adapted to provide an analog voltage in proportion to the number of digital signals 1 in each of the pattern component signals. The pattern component extracting unit 4 of the type described is well known in the art. The output signals of the pattern component extracting unit 4 are transferred through output lines 4.sub.1 to 4.sub. n to a similarity detecting unit 5 and to an auxiliary identification unit 9.

The n analog voltages transferred into the similarity detecting unit 5 are compared with a number of m standard patterns in storage and converted into a number of m analog voltages each representing the degree of similarity to each of the standard patterns. The m analog voltages are transferred through output lines 5.sub.1 to 5.sub.m to a maximum and next-maximum similarity detecting unit 6, which is adapted to convert the maximum analog voltage among said m analog voltages and the analog voltages whose ratio to said maximum analog voltage is higher than a predetermined ratio, for instance, 70 percent, into the digital signals "1"s and to convert the analog voltages whose ratio is less than 70 percent into the digital signals "0"s. The output signals of the detecting unit 6 are transferred through output lines 6.sub.1 to 6.sub.m to an input pattern identification unit 7.

The identification unit 7 is adapted to provide the digital 1 signal on one of output lines 7.sub.1 to 7.sub.m, each of which corresponds to each of m standard patterns of different categories, in response to a timing pulse CP.sub.1 transmitted through an input line 7B when only one of the m output signals transmitted from the detecting unit 6 is 1. The identification unit 7 is adapted to provide a strobe signal STB on an output line 7C simultaneously when it provides the digital 1 signal on one of the output lines 7.sub.1 to 7.sub.m. When the digital 1 output signals transmitted from the detecting unit 6 are three or more, the identification unit 7 provides the strobe signal STR in response to the timing pulse CP.sub.1, but provides the digital 0 signals on all of the output lines 7.sub.1 to 7.sub.m. This means that the identification unit 7 cannot identify the input pattern information. When the detecting unit 6 provides two digital 1 output signals, the identification unit 7 provides not only two digital 1 signals on corresponding two of the output lines 7.sub.1 to 7.sub.m but also a pattern enlargement signal in response to the timing signal CP.sub.1.

A pattern enlargement control unit 8 is adapted to enlarge a specific region of the input pattern information stored in the memory 3 in response to the pattern enlargement signal transmitted from the identification unit 7. The enlarged specific region is converted into n analog voltages by the pattern component extracting unit 4 and transferred into the auxiliary identification unit 9. It should be noted that two digital 1 output signals of the detecting unit 6 are maintained even when the specific region is enlarged.

The auxiliary identification unit 9 comprises q/2 adder-substractors each of which is adapted to identify to which of the two standard patterns the enlarged input pattern belongs. The identification unit 7 is adapted to provide the digital 1 output signal on one of the output lines 7.sub.1 to 7.sub.m in response to the two digital 1 output signals from the detecting unit 6 and in response to the output signal transmitted to the identification unit 7 from the auxiliary identification unit 9 through one of its output lines 9.sub.1 to 9.sub.q. In this case the digital 1 output signal is provided in synchronism with the timing pulse CP.sub.2 transmitted through an input line 7B.sub.2, and the strobe signal STB is also simultaneously provided. However, when no output signal is transmitted from the auxiliary unit 9 to the identification unit 7, the latter provides only the strobe signal STB in synchronism with the timing pulse CP.sub.2. This means that the identification of the input pattern is impossible.

FIG. 2 shows an example of a pattern, that is a numeral "5" stored in the two-dimensional memory 3. When this pattern is identified as "5" and "6" as the result of the similarity detection by the unit 5, a specific region 3A of the pattern shown in FIG. 5(a), which is determined in response to the control signals transmitted from the identification unit 7 through output line 7A.sub.1 to 7A.sub.p, is enlarged, for instance, two times as shown in FIG. 5(b). In general the two-dimensional memory 3 is made up by a plurality of shift registers so that any partial region of the pattern stored in the memory may be easily enlarged based upon the conventional digital techniques.

Next referring to FIG. 3, the identification unit 7 will be described in more detail hereinafter. The outputs of the maximum and next maximum detecting unit 6 (See FIG. 6) are applied through the output lines 6.sub.1 to 6.sub.m to an adder circuit comprising resistors RB.sub.1 to RB.sub.m (where RB.sub.1 = RB.sub.2 = RB.sub.m) and RA.sub.1 and an operational amplifier 71. The reference voltage V.sub.1 is so selected that when the detecting unit 6 provides three or more digital 1 signals, a first comparator 72-1 provides the digital signal 1 while the reference voltage V.sub.2 is so selected that when the detecting unit 6 provides two or more digital 1 output signals, a second comparator 72-2 provides the digital 1 output signal. When the digital 1 output signal appears, for instance, only on the output line 6.sub.ml, the outputs of the both first and second comparators 72-1 and 72-2 are "0"s so that the outputs of the inverters 73-1 and 73-2 are "1"s. Therefore an AND gate 74-9 is opened in response to the outputs "1"s of the inverters 73-1 and 73-2 and to the timing pulse CP.sub.1 so that the strobe signal STB is transmitted through an OR gate 75. Furthermore the identification signal is transmitted through an AND gate 74-1 and an OR gate 76-1 to the output line 7.sub.ml. When the detecting unit 6 provides three or more digital "1" output signals, an AND gate 74-7 is opened in response to the output of the comparator 72-1 and the pulse CP.sub.1 so that the strobe signal STB also appears on the output line 7C of the OR gate 75, but no signal appears on the output lines 7.sub.1 to 7.sub.m. When the digital 1 output signals appear on the output lines 6.sub.ml and 6.sub.m2, an AND gate 74-8 is opened in response to the output of the inverter 73-1, the output of the comparator 72-2 and the pulse CP.sub.1 to transmit the signal to the input of an AND gate 77 to which is also applied the digital 1 signal through an AND gate 74-5 and an OR gate 76-3. Therefore the AND gate 77 is opened to transmit the enlargement signal through the output line 7.sub.pl, so that the pattern enlargement control unit 8 (See FIG. 1) is actuated to enlarge a specific region of the input pattern stored in the two-dimensional memory 3. Thus the outputs of the auxiliary identification unit 9 are applied to the identification unit 7 through the output lines 9.sub.1 to 9.sub.q.

It is now assumed that the categories cm1 and cm2 correspond to the output signals which appear on the output lines 6.sub.m1 and 6.sub.m2, respectively and that the output signals which are used for identification of the categories cm1 and cm2 are transmitted through the output lines 9.sub.q1 and 9.sub.q2, respectively. That is, the output signal 1 on the output line 9.sub.q1 designates the category cm1 while the digital 1 output signal on the output line 9.sub.q2 denotes the category cm2. On the other hand, when the digital 0 signals appear on both output lines 9.sub.q1 and 9.sub.q2, the identification of the input pattern becomes impossible. When the digital 1 output signal appears on the output line 9.sub.q2, the OR gate 75 provides the strobe signal STB in response to the timing pulse CP.sub.2, and an AND gate 74-4 is opened in response to the output signals transmitted through the output lines 6.sub.m1, 6.sub.m2 and 9.sub.q2 so that the digital 1 signal appears on the output line 7.sub.m2 of an OR gate 76-2. Thus, the input pattern is identified as belonging to the category cm2.

FIG. 4 is a fragmentary circuit diagram of the auxiliary identification unit 9. The n analog output voltages from the pattern component extracting unit 4 are transmitted through the output lines 4.sub.1 to 4.sub.n to an adder-subtractor comprising resistors R.sub.1 to R.sub.n, RA.sub.2 and RA.sub.3 and an operational amplifier 91 which is connected to a first and second comparators 92-1 and 92-2. When the enlarged input pattern belongs to the category cm1, the digital 1 output signal appears on the output line 9.sub.q1. In like manner, when the enlarge input pattern is identified as belonging to the category cm2, the digital 1 output signal appears on the output line 9.sub.q2. The reference voltage V.sub.4 is so selected as to be higher than the reference voltage V.sub.3. The values of the resistors R.sub.1 to R.sub.n and whether they are connected to the positive or negative polarity input terminals of the operational amplifier 91 are so determined that the enlarged input pattern may be identified in the most stable manner. The values of the resistors RA.sub.2 and RA.sub.3 are selected depending upon those of the resistors R.sub.1 to R.sub.n.