Neural Network System for Image Processing

1. Field of the Invention

The present invention relates to a visual information processing device which is capable of processing speed information (optical flow) for an object moving at high speed.

2. Description of the Related Art

FIG. 1 shows a three-dimensional resistance circuit network 1 for processing an optical flow, as presented, for example, in ["Neural Computer", Berlin, (1988), p.101, C. Koch, et al]. Shown in the figure are a variable power supply 2, a capacitor 3, a variable resistor 4, connecting points 5 and 6 of the three-dimensional resistance circuit network 1, a resistor 7 for connecting the points of an upper layer and an under layer, a variable resistor 8 for connecting the points in an upper layer surface (or an under layer surface), and a switch 9 connected to the resistor 8.

Next, the operation will be explained. The voltage of the variable power supply 2, the value of the variable resistor 4, and the value of the variable resistor 8 for connecting the points, which are determined from input data obtained from photoelectric conversion devices such as photodiodes, are simultaneously input to the connecting points of the three-dimensional resistance circuit network at a certain time interval. The variations in the voltage at the connecting points 5 and 6 resulting from a variation in an optical signal, caused by the movement of an object, disturbs the circuit network system. The three-dimensional resistance circuit network then enters a new, stable state of minimum power consumption in accordance with Kirchoff's law. From the voltage value of each of the connecting points 5 and 6 which have reached a stable state, two-dimensional speed vectors in the x and y directions can be determined. In this case, however, since the contour of the object becomes unclear, the switch 9 is provided so that a contour can be formed at a rational position.

A conventional visual information processing device has been constructed as mentioned above. As a result, the following problems exist in that the construction of a three-dimensional resistance network in an integrated circuit is difficult and the formation of photoelectric conversion devices such as CCDs, photodiode arrays or the like in the same information processing device is also difficult.

The present invention has been devised to eliminate the above problems. It is an object of the present invention to provide a visual information processing device which can be efficiently incorporated in other devices and which is capable of processing the speed information for a moving object in realtime.

A visual information processing device according to the present invention comprises a pair of layered neural network units each having an integrated circuit device in which a neuron circuit region is provided in a matrix form and a molecule film device provided on the integrated circuit region. The molecule film device has a photoelectric function section having a construction in which a molecule multilayer having a heterojunction is sandwiched between electrodes. The photoelectric function section includes a video signal input section in which one pixel corresponds to one neuron and a T.sub.ij signal input section. The neuron circuit region has a structure in which the four closest neurons are connected in the same layer and the one-to-one corresponding neurons are connected between the layers.

The visual information processing device of the present invention is constructed as mentioned above. A bonding strength matrix T.sub.ij between neurons can be written from the outside and, what is more, a video signal can be input simultaneously. Neurons are bonded so that the speed information on a moving object can be obtained by an interconnection only between closest neurons. As a result, the number of interconnections among neurons and of T.sub.ij can be diminished and input video information can be processed at high speed.

According to the present invention, a video signal input section and a T.sub.ij signal input section are provided in a photoelectric function section formed by a molecule multiplayer having a heterojunction sandwiched between electrodes. A video input signal has a one-pixel one-neuron correspondence. A neural network, in which a network is assembled so that the respective neurons are connected with the five neighboring neurons in the T.sub.ij signal input section of the photoelectric function section, is constructed on an integrated circuit having only the neuron sections incorporated. Therefore, a T.sub.ij signal can be written from the outside at the same time an image sensor starts to function, thus enabling a visual information processing system capable of processing the video signal input and the speed vector of a moving object at a high speed to be obtained.

Also, according to the present invention, since a photoelectric function section having a construction in which a molecule multilayer having a heterojunction is sandwiched between electrodes can be easily formed on an integrated circuit, the density of neurons per unit area in an information processing device can be easily increased by constructing a three-dimensional visual information processing device.

These and other objects, features and advantages of the present invention will become clear when reference is made to the following description of the preferred embodiments of the present invention, together with reference to the accompanying drawings.

FIG. 1 is a conventional three-dimensional resistance circuit network for calculating the speed information of an object;

FIGS. 2(a)-(c) are views illustrating the placement and interconnection of an upper-layer neural network of a visual information processing device of the present invention;

FIGS. 3(a)-(b) are configurational views of a visual information processing system constructed using the visual information processing device of FIG. 2;

FIG. 4 is a schematic illustration of a neuron network of the visual information processing device of the present invention;

FIG. 5 is a simplified conceptual diagram of the neural network of FIG. 2(a); and

FIG. 6 is an electrical schematic of FIG. 5.

An embodiment of the present invention will be explained hereinunder with reference to the accompanying drawings.

FIG. 2 is an illustration of a visual information processing device according to one embodiment of the present invention. In this figure, only one layer of a two-layer structure is shown. Shown in the figure are a neuron circuit region 11 (hereinafter abbreviated to a neuron) formed in an integrated circuit, an input terminal 12 of the neuron, an output terminal 13 of the neuron, a molecule multilayer 16 having a heterojunction, an input wiring 14, under the molecule multiplayer 16, on a semiconductor integrated circuit device, an output wiring 15 on the molecule multilayer 16, an insulation film 17 formed on the integrated circuit, a video signal input section 18a, a silicon substrate 19 forming the integrated circuit, a cross section 20 of the neuron region formed in the integrated circuit, and a section 18b for inputting an T.sub.ij signal representing the input and output synapse bonding strength of neighboring neurons. (FIG. 5 is a simplified conceptual diagram of the neural network of FIG. 2A). FIG. 6 is an electrical schematic diagram of FIG. 5. In FIGS. 5 and 6, V.sub.A. . . V.sub.D correspond to video signal input sections 18(a) of FIG. 2(a), while N.sub.A. . . N.sub.D correspond to neuron circuits 11 of FIG. 2(a). Also in FIGS. 5 and 6, sections T.sub.AB, T.sub.BA. . . T.sub.DC, T.sub.CD correspond to sections 18b of FIG. 2(a).

The input wiring 12 is formed by vacuum deposition of aluminum and photo-lithography, and the output wiring 13 is formed as a semitransparent aluminum electrode by vacuum deposition of aluminum and photo-lithography. As for the output wiring 13, a transparent electrode such as ITO, Nesa glass or the like can be used. The heterojunction molecule multiplayer 16 is formed in such a way that several layers of hematoporphyrin(IX)-bis (tridecanoylether):Ru(P(OCH.sub.3)3)2 [abbreviated to RuHP(Ph).sub.2 ] are deposited on an integrated circuit on which an under-layer aluminum electrode is formed, by the Languir-Blodgett:LB method and several layers of 7,8-dimethyl-3,10-dinonyl isoalloxiazine (abbreviated to DNI) are deposited thereon. Such a molecule multiplayer having a heterojunction has such properties that a current flowing through electrodes varies according to the intensity of irradiated light.

The neurons formed in the integrated circuit are formed in combination with resistors, capacitors, transistors and so forth by using a conventional LSI fabrication technology.

FIG. 3 shows a visual information processing system constructed using the visual information processing devices shown in FIG. 2. Shown in FIG. 3 are an input video signal 21, a device 22 for emitting a T.sub.ij pattern, a half-mirror 23 that transmits an input image signal as it is and that changes the T.sub.ij signal in the same direction as that of the input image signal, color filters 24a and 24b formed on the heterojunction molecule multilayer 1, a color filter 25 for converting the input video signal to a monochrome signal, an optical lens system 26 for focusing the input video signal and the T.sub.ij signal on a photoelectric conversion section of the visual information processing device, the visual information processing device 27 mentioned above, and an output signal 28.

FIG. 4 is a schematic illustration of a neuron network of the visual information processing device. In this figure, the reference number 30 denotes a simplified wiring showing the bonding among the neurons.

Next, the operation will be explained.

The input video signal 21 mixed with a static object and a moving object is irradiated through the band pass filter 25 of 450 nm and regulated by the optical lens system 26 so that the video signal is focused on the video input section 18a of the heterojunction molecule multilayer 16. The filter 24a capable of transmitting light having a wavelength of 500 nm or less is disposed on the video input section 18a.

The T.sub.ij optical signal having a wavelength of about 550 to 600 nm is irradiated from the T.sub.ij matrix light-emitting device 22, the light path of which is bent by the half-mirror 23, and is irradiated on the T.sub.ij input section 18b of the heterojunction molecule multilayer 16. The filter 24b that does not transmit light having a wavelength of 500 nm or less is attached on the T.sub.ij input section 18b and this section does not respond to the input video signal.

In the integrated circuit device, the neurons 20 form a network having a two-layer matrix structure shown in FIG. 4 in such a manner that the input and output of neighboring neurons are bonded one another. FIG. 2 shows the wiring state of the structure of one layer of the two-layer structure. Each of the bonding lines 30 among the neurons shown in FIG. 4 is represented by the corresponding two lines in the embodiment in FIG. 2. The bonding section of the input and output of the neighboring neurons corresponds to the T.sub.ij input section 18b of the heterojunction molecule multilayer 16. The strength among the respective input/output bonding sections (actually, the magnitude of the resistance thereamong) is determined from the intensity of the light which is irradiated from the respective light matrix elements in the T.sub.ij matrix light-emitting device 22 onto the respective input/output bonding sections.

The video input signal is input to the neural network at a certain time interval. If the first video input signal is simultaneously applied to the respective video input sections 18, the input states of the respective neurons are changed so that the entire neural network instantly enters a stable state. When the second video input signal is applied to the sections 18 after a certain time interval, (i) if there is no change between the first and second video input signals, the entire neural network remains in a stable state, and on the other hand (ii) if any video information representative of a moving object is included in the second video input signal and such video information is offset from that included in the first video input signal, although the input states of the respective neurons are changed to disturb the stable state of the neural network, it then instantly enters the next stable state. In the newly entered stable state, the speed vector of the moving object in an X direction can be determined from the output values of the respective neurons of the upper layer, and the speed vector in a Y direction can be determined from the output values of the respective neurons of the lower layer. From these speed vectors, the information regarding the position and speed of the object which has moved can be obtained. In FIG. 3, the images of both a static flower and a flying bee are shown as the input video signal 21, and the image of the flying bee is extracted as the output video signal 28 from the visual information processing device 27.