JPH07160796A - Signal processor by neural network - Google Patents

Abstract

PURPOSE:To execute signal processing (enhancement processing, noise reduction processing) by composing the neural network of limited precision, a small scale and excellent learning efficiency. CONSTITUTION:The object of the composing of the neural network (1) is set to be an existent signal processor itself which processes a signal after deterioration to be that before deterioration and the object of the learning of the neural network is set to be data obtained by processing random data by the existent signal processor.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a signal processing device (for example, an image enhancing / image noise reducing device) based on a neural network, and more particularly, it has a simpler structure than that of an existing signal processing device, and it is practical. The present invention relates to a signal processing device using a neural network, capable of constructing a neural network.

[0002]

2. Description of the Related Art Conventionally, there are image enhancing devices and image noise reducing devices that use a linear filter or a non-linear filter of an analog circuit. In the enhancement process, a high frequency component is emphasized by a filter in order to emphasize the contour of the image. In the noise reduction process, the high frequency component is removed by a filter to reduce noise. As described above, in an image processing circuit using an analog circuit, it is necessary to design a filter device suitable for processing characteristics and manufacture and adjust the device.

There is also a digital image processing apparatus for performing image processing by calculating digital data by a digital circuit. For example, as shown in FIG.
Enhancement processing and noise reduction processing were performed based on the image data of the location. Point Y is a point to be processed, and peripheral points X1 to X8 are signals before and after a predetermined pixel period and horizontal line period, respectively. Here, the brightness data (level) of the point Y is y, and the brightness data (level) of the surrounding points X1 to X8.
And the deviation m is m = 9y−S. At this time, if the luminance data (level) of the point Y is set to S / 9 and the luminance data (level) of the surrounding points X1 to X8 is averaged, noise reduction processing is performed. If the luminance data (level) of the point Y is set to y + 0.3 m and emphasis is made according to the surrounding points, the enhancement processing is performed. Further, when the magnitude of the deviation m is less than a certain value, the luminance data (level) of the point Y
Is S / 9, and when the magnitude of the deviation m is a certain value or more,
If the luminance data (level) at the point Y is set to y + 0.3 m, it also serves as a device for performing noise reduction processing and enhancement processing. An image processing apparatus based on digital calculation requires a large-scale digital signal processing apparatus for constructing an algorithm suitable for processing characteristics and executing the algorithm.

As described above, in the existing signal processing apparatus (the filter of the analog circuit or the digital image processing apparatus described above), as shown in FIGS. 3A and 3B, when signal deterioration occurs in the transmission system. The signal processing device manufactured and adjusted processed the deteriorated signal and output it. On the other hand, recently, development of a signal processing device using a neural network has been attempted. If a neural network is used, learning processing is required, but unlike the existing signal processing device described above, analysis of complicated processing procedures (algorithms) and manufacturing / adjustment of devices for realizing it are unnecessary. is there.

As shown in FIG. 4, in a signal processing device using a neural network, a neural network is preliminarily constructed by performing a learning process from an original signal before transmission (before deterioration) and a transmitted signal after deterioration. To be done. Before transmission (before deterioration)
The original signal of is a teacher signal, the transmitted signal after deterioration is used as a signal for learning, and when the signal used for learning is input to the neural network, the difference between the output of the neural network and the teacher signal becomes small, The weight between the units is calculated, and the neural net is constructed. Then, when a deteriorated signal is input to the neural network constructed by the learning process, a signal similar to the original signal before deterioration is output.

When the image signal is processed by the neural network, image signal data such as a natural image and a test pattern is used as the original signal before transmission (before deterioration). Then, as the transmitted signal after deterioration, for example, in the case of a signal processing device (enhancement or image noise reduction) that copes with image deterioration due to recording / reproduction on a VTR (video tape recorder), it is actually recorded by the VTR. An image that has been reproduced and deteriorated (image signal data) is used.

[0007]

However, in the conventional signal processing apparatus using the neural network, the neural network must be enlarged to some extent (in the case of processing an image signal by the three-layer neural network shown in FIG. If you do not increase the number of units in the layer), the learning process does not work well and the weight calculation process diverges, or the number of times for learning becomes very large, and it is possible to construct a realistic neural network. There were things I couldn't do.

Further, when learning is performed using image signal data such as a natural image, the processing effect (of the constructed neural network) is good for a specific image, but the processing effect is extremely poor because the learning target is biased. There was a chance Further, when learning is performed using image signal data such as a natural image, the weight calculation process may diverge without being learned, and the number of times of learning may be extremely large.

Therefore, the present invention stops construction of a high-precision and large-scale neural network like a conventional neural network, and provides a small-scale neural network having only required accuracy and excellent learning efficiency. It was constructed and tried for signal processing.

[0010]

In order to solve the above problems, the present invention processes a signal after deterioration into a signal before deterioration (for example, noise reduction processing and In a signal processing apparatus using a neural network (1) for enhancing processing, an existing signal processing apparatus itself (for example, noise reduction processing) that processes a signal after deterioration into a signal before deterioration, in a signal processing apparatus using the neural network (1). And an image processing device by digital calculation for enhancing processing), and a learning target of the neural network is random data processed by the existing signal processing device. To provide a signal processing device according to the present invention.

[0011]

According to the signal processing device by the neural network configured as described above, the target of construction of the neural network is not the direct inverse characteristic of the system in which signal deterioration occurs, but the target of construction is the existing signal. It is limited to the processor (system). Furthermore, the target of learning of the neural network is not limited to biased data such as natural images and test patterns,
It will be a high range.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of a signal processing device using a neural network according to the present invention will be described in detail below with reference to the accompanying drawings. As described above, the signal processing device by the neural network according to the present invention, like the conventional neural network, stops the construction of a high-precision and large-scale neural network, and has a small scale equipped with only the required accuracy. This is an attempt to perform signal processing by constructing a neural network with excellent learning efficiency and has the following features.

First, the target of the neural network construction is not the direct inverse characteristic of the system in which signal deterioration occurs, that is, the signals before and after the system in which signal deterioration occurs as shown in FIG. The object of learning is not the signal used for learning but the processing itself executed by the existing signal processing apparatus, that is, the signals before and after the processing executed by the conventional apparatus as shown in FIG. Signal Used as a signal for learning. Second, the learning target (learning material) of the neural network is not random data that actually exists (for example, a natural image or a test pattern) but mechanically generated random data.

Here, the signal processing apparatus based on the present neural network will be specifically described by taking an apparatus that performs the enhancement processing and the noise reduction processing as an example. As shown in FIG. 5 described above, the present apparatus performs the enhancement processing and the noise reduction processing based on the image data of nine places that are temporally before and after. First, the configuration of a signal processing device using a neural network will be described. As shown in FIG. 2, the input image data (for example, luminance signal data) is delayed by the delay element 2 having a delay amount of one horizontal line period and the delay element 3 having a delay amount of one pixel period. Then, at the position (time) interval shown in FIG. 5, the target pixel Y and the peripheral pixels X1 to X8 which are temporally and spatially different from the target pixel Y are neural images as image data at a total of nine locations. It is configured to input to the net 1.

The neural network 1 has a layered structure (for example, an input layer, a backpropagation learning method).
It is a well-known neural network of three layers (intermediate layer and output layer), and is a combination of a large number of units (neurons, shown by large white circles) connected with variable weights. The unit is a conversion system with non-linear input / output characteristics, and the unit has a weight independent of the output value from the previous layer (indicated by a small black circle)
The sum total from the front layer obtained by multiplying by is input, and the sum is nonlinearly converted and output to the rear layer. Input layer is 9 units, middle layer is 4 units, output layer is 1 unit
It is composed of individual units.

In learning of the neural network 1,
As shown in FIG. 1, the process itself (algorithm) executed in an existing signal processing device (for example, an image processing device by digital calculation that performs the noise reduction process and the enhancement process described above) is targeted for construction. In other words, the signals used before and after learning are processed by the conventional device.
Use as a teacher signal.

As described above, the existing digital image processing apparatus performs noise reduction processing and enhancement processing by the following algorithm processing. The brightness data (level) of the processing target point Y is y, and the surrounding points X1 to X8.
Of the brightness data (level) of S, the deviation m is m = 9y
As -S, the brightness data (level) of the point Y is set to S / 9 when the magnitude of the deviation m is equal to or less than a fixed value, and the brightness data (level) of the point Y is set to (S) when the magnitude of the deviation m is equal to or more than the fixed value. level)
Is y + 0.3m.

Therefore, in the learning of the neural network 1, the luminance data (level) of the point Y to be processed and the peripheral points X1 to X8 is used as a signal for learning, and the teacher signal at this time is calculated by the above algorithm. Value. Backpropagation is used as a learning method so that the weight between units is varied and learning is performed so that the difference between the output value of the neural network 1 and the teacher signal becomes small.

At this time, the brightness data (the brightness data of the processing target point Y and the peripheral points X1 to X8) which is the learning target (learning material) of the neural network actually exists (for example, naturally). Image data consisting of mechanically generated random data is input instead of image data or test pattern data. This random data is used as a signal used for learning, and the value calculated by the above algorithm based on this random data is used as the teacher signal.

Note that generation of image data (signals used for learning) consisting of random data, calculation of a teacher signal based on an algorithm similar to the processing executed by an existing device, variable convergence processing of weights by backpropagation, etc. , Is processed on a workstation (not shown).

Table 1 compares the learning process and the image processing result by the conventional neural network signal processing apparatus and the neural network signal processing apparatus according to the present invention. In Conventional Examples 1 and 2, the target of the neural network construction is the direct inverse characteristic of the system in which signal deterioration occurs (that is, the signals before and after the system in which signal deterioration occurs are used as teacher signals / learning signals). Yes, the target of the neural network learning is the data of the natural image. Conventional Example 1 is a relatively large-scale neural network having eight units in the middle layer. Conventional example 2 is a relatively small-scale neural network having four units in the intermediate layer. In addition,
As described above, the neural network of the signal processing device according to the present invention has a relatively small number of units because the number of units in the intermediate layer is four. In the conventional examples 1 and 2, except the target of the neural network construction, the target of learning, and the number of units in the intermediate layer,
This is the same as the signal processing device using the neural network shown in FIG.

[0022]

[Table 1]

As is clear from Table 1, according to the signal processing device by the neural network according to the present invention, the object of construction is limited, so that the small scale (units of the intermediate layer) having only the required accuracy is provided. It is possible to construct a neural net with excellent learning efficiency. In other words, as in Conventional Example 1, because of high accuracy, the scale is large (the number of units in the middle layer is about 8) and the number of learning increases, and conversely, as in Conventional Example 2, small scale (units in the middle layer). Since the number is about 4), construction (learning) does not become impossible.

Further, according to the signal processing device by the neural network according to the present invention, since the object of learning is set to a high range such as random data so that no bias is generated, an appropriate processing effect (levels 4 to 5) is obtained. ) Can be obtained by moderate learning. That is, the processing effect is very good for the image close to the natural image used for learning as in the conventional example 1 (level 7 to 7).
However, if the natural image used for learning is significantly different from 8), the processing effect will not be extremely bad (level 2 to 2).
3).

Further, the neural network in which the number of units in the intermediate layer is about 4 can be composed of operational amplifiers, resistors and the like, and can be composed of a small number of parts as compared with existing filters and digital signal processing devices.

Therefore, it is possible to construct a realistic neural network with a simpler configuration than the existing signal processing apparatus, and it is possible to perform signal processing using the neural network.

In this embodiment, the existing device for constructing the neural network is an image processing algorithm (digital signal processing device) by digital calculation for performing the noise reduction process and the enhancement process described above. An image enhancing device or an image noise reducing device using a linear filter or a non-linear filter with an analog circuit may be a target for constructing a neural network, and an input / output signal of the filter may be a signal / teaching signal used for learning. That is, the signals before and after being processed by the existing device may be used as the signals / teaching signals used for learning.

In the present embodiment, the luminance signal (luminance data) has been described as an example, but the same effect can be obtained by implementing the RGB color signal or the color difference signal. The processing content is not limited to the noise reduction processing and the enhancement processing.

[0029]

As described above, the signal processing device using the neural network according to the present invention is an existing signal processing device for processing a signal after deterioration into a signal before deterioration as a target for constructing a neural network. In addition to itself, since the learning target of the neural network is the data obtained by processing random data with the existing signal processing device, the target of construction is limited, and the learning target is high without being biased. It is possible to construct a neural network with a high learning efficiency on a small scale with only the required accuracy. Therefore, it is possible to construct a realistic neural network with a simpler configuration than the existing signal processing apparatus, and it is possible to perform signal processing using the neural network.

[Brief description of drawings]

FIG. 1 is a principle diagram illustrating an embodiment of a signal processing device using a neural network according to the present invention.

FIG. 2 is a specific configuration diagram illustrating a signal processing device using a neural network according to the present invention.

FIG. 3 is a diagram illustrating a concept of signal processing by an existing signal processing device.

FIG. 4 is a principle diagram illustrating a conventional signal processing device using a neural network.

FIG. 5 is a diagram showing image data at nine positions before and after in terms of time (position) for explaining the signal processing.

[Explanation of symbols]

 1 Neural network 2,3 Delay element X1 to X8 Peripheral points Y Point to be processed

Claims (2)

[Claims] 1. A signal processing apparatus using a neural network for processing a deteriorated signal into a signal before deterioration, wherein an existing target for constructing the neural network is an existing processor for processing a signal after deterioration into a signal before deterioration. A signal processing device by a neural network, wherein the signal processing device itself is used, and the learning target of the neural network is data obtained by processing random data by the existing signal processing device. 2. A signal processing device using a neural network according to claim 1, wherein an enhancement process or a noise reduction process is performed.