JPH03225482A - Neural network expressing position, size, and both those of target object - Google Patents

Abstract

PURPOSE:To perform robust and accurate pattern recognition for the superposition of a noise, the deviation of a position, and the deviation of size by providing a position network and a size network which detect the rough position and size of a pattern with arbitrary shape and size that exists on a two-dimensional screen, respectively. CONSTITUTION:At the position network, forty neuron units in an intermediate layer are connected to a neuron unit in an input layer in network shape, and are connected to two neuron network in an output layer conformed to an x-y orthogonal coordinate in the network shape. The position network is the one to detect the position of the pattern, for example, center or gravity (x.y). Also, in the size network, the forty neuron units in the intermediate layer are con nected to the neuron unit in the input layer in the network shape, and also, is connected to one neuron unit in the output layer conformed to one side of a square circumscribed to a target pattern. The size network detects the size of the pattern, for example, the length (d) of one side of the square enclosing the pattern. In such a way, the position and the size of the target pattern can be detected.

Description

DETAILED DESCRIPTION OF THE INVENTION [Purpose] [Industrial Field of Application] The present invention relates to a neural network that is used for image recognition, pattern recognition, etc., and expresses the position, size, and both of a target object.

[Prior Art] In recent years, research on neural networks that imitate the brain has become active, and applications are being attempted in various fields. A pack propagation model (hereinafter referred to as BP model) with a hierarchical structure, which is a neural network model.
Due to its learning ability and nonlinear function approximation ability, attempts have been made to apply it to pattern recognition of handwritten characters, etc. Neural networks based on conventional BP models are strong against deformation and noise of the pattern to be recognized, that is, they can correct correctly, but they hardly take into account positional and size deviations from the learned pattern. It has not been. That is, in the conventional BP model, it has not been possible to eliminate positional deviation and size deviation.

[Problem 3 to be Solved by the Invention As described above, the BP model has the problem that the position and size cannot be detected correctly depending on the state of the target pattern.

An object of the present invention is to provide a neural network that can accurately detect the position, size, and both of a target pattern within a two-dimensional field of view.

“Configuration” [Means for solving the problem] The neural network for expressing the position of a target object according to the present invention uses an input divided into a grid in a BP model that defines connections between neurons and a learning rule. a neuron unit in an input layer that inputs a two-dimensional image as an intensity signal of each grid; a neuron unit in an output layer that outputs the position of a target pattern in the input two-dimensional image; a neuron unit in the input layer; and an intermediate layer neuron unit provided between output layer neuron units, and is characterized in that it detects the position of the pattern having an arbitrary shape and size existing on the input two-dimensional image.

The neural network that expresses the size of a target object according to the present invention uses an input two-dimensional image divided into grids as an intensity signal of each grid in a BP model that defines connections between neurons and learning rules. A neuron unit in the input layer that receives input, a neuron unit in the output layer that outputs the size of the target pattern of the input two-dimensional image, and a neuron unit provided between the neuron unit in the input layer and the neuron unit in the output layer. The present invention is characterized by detecting the size of the pattern having an arbitrary shape and size existing on the input two-dimensional image.

The neural network that expresses the position and size of a target object according to the present invention uses an input two-dimensional image divided into grids as an intensity signal for each grid in a BP model that defines connections between neurons and learning rules. The first input as
a neuron unit in the input layer, a neuron unit in the first output layer that outputs the size of the target pattern of the input two-dimensional image, and a neuron unit in the first input layer and a neuron unit in the first output layer; a first intermediate layer neuron unit provided in the input two-dimensional image, and a position net for detecting the size of the pattern having an arbitrary shape and size existing on the input two-dimensional image; a second input layer neuron unit that inputs the input two-dimensional image as an intensity signal of each grid; a second output layer neuron unit that outputs the size of the target pattern of the input two-dimensional image; The pattern is composed of neuron units of one input layer and neuron units of a second intermediate layer provided between the neuron units of the second output layer, and has an arbitrary shape and size existing on the input two-dimensional image. and a size net that detects the size of.

The neural network that expresses the position and size of a target object according to the present invention uses an input two-dimensional image divided into grids as an intensity signal for each grid in a BP model that defines connections between neurons and learning rules. The first input as
a neuron unit in the input layer, a neuron unit in the first output layer that outputs the size of the target pattern of the input two-dimensional image, and a neuron unit in the first input layer and a neuron unit in the first output layer; a first intermediate layer neuron unit provided in the input two-dimensional image, and a position net for detecting the size of the pattern having an arbitrary shape and size existing on the input two-dimensional image; a second input layer neuron unit that inputs the input two-dimensional image as an intensity signal of each grid; a second output layer neuron unit that outputs the size of the target pattern of the input two-dimensional image; The pattern is composed of neuron units of two input layers and neuron units of a second intermediate layer provided between the neuron units of the second output layer, and has an arbitrary shape and size existing on the input two-dimensional image. a size net that detects the size of the image, a neuron unit in the third input layer that inputs the input two-dimensional image divided into grids as intensity signals for each grid, and neurons corresponding to four directions (up, down, left, and right). A neuron unit of the third output layer, a neuron unit of the third input layer, and a neuron unit of the third input layer whose teacher value is 0 when moving the frame inward relative to the unit and 1 when moving the frame outward. and a third intermediate layer neuron unit provided between the neuron units of the layer,
[Function] Position, size, Framing has been considered as a normalization process in the preprocessing of feature extraction in conventional pattern recognition, but both assume sequential processing of address specification memory. A major feature of the present invention is that these preprocessings are also considered as pattern conversion problems based on parallel processing, and are realized as network learning.

Due to this feature, by devising the learning pattern to deal with deformation and noise, it is possible to approach human recognition, and the position and size of the target pattern can be detected accurately.

That is, with the position net of the present invention, the problem of detecting the position of a pattern within the field of view can be solved as pattern conversion using the BP model, and with the size net of the present invention,
BP solves the problem of detecting the size of a pattern within the field of view.
Since it can be solved as a pattern transformation using a model, the position and size of the target pattern can be detected accurately.

Note that the framing net used in the present invention can set an approximate frame circumscribing the pattern, correct the generated frame, and narrow down the range of the pattern. The ability to correct the size can be acquired through learning, and the effects of noise can be removed.

[Example] Referring to FIG. 1, an example of the location net of the present invention will be described. FIG. 1(a) shows an input two-dimensional image,
It is divided into 6 parts in the X direction and 6 parts in the X direction, for a total of 36 grids. The input signals from each grid are input to the 36 input layer neuron units shown in FIG. 1(b). The 40 neuron units in the intermediate layer are connected in a network to the neuron units in the input layer, and are connected in a network to two neuron units in the output layer that correspond to xy orthogonal coordinates. This position net is the position of the pattern, e.g. centroid (x
, y).

An embodiment of the size net of the present invention will be described with reference to FIG. Figure 2(a) shows the input two-dimensional image, which is divided into 6 parts in the X direction and 6 parts in the X direction, for a total of 36 parts.
divided into grids.

The input signal from each grid is 36 as shown in FIG. 2(b).
input to the neuron units of the input layer. The 40 neuron units of the intermediate layer are connected in a network to the neuron units of the input layer, and are connected to one neuron unit of the output layer corresponding to a piece of a square circumscribed to the target pattern. This size net detects the size of the pattern, for example, the length d of one side of the square surrounding the pattern.

An embodiment of the framing net of the present invention will be described with reference to FIG. Figure 3(a) shows the input two-dimensional image, which is divided into 6 parts in the X direction and 6 parts in the X direction, for a total of 36 parts.
divided into grids.

The input signal from each grid is 36 as shown in FIG. 3(b).
input to the neuron units of the input layer. The 40 neuron units in the intermediate layer are connected to the input layer neuron units in a network, and the 40 neuron units in the upper, lower, left, and right
It is connected to four neuron units of the output layer that correspond to the directions.

This frame cutting net outputs the component in the direction in which the frame is thought to protrude as 1 in the four directions of the top, bottom, left, and right, and the other components as 0. When the value is 1, the frame is moved inward, and when it is 0, the frame is moved outside.

Referring to FIG. 4, a pattern recognition architecture applying position nets and size nets will be described. As shown in FIG. 4, the pattern recognition architecture of the present invention is constructed around four networks: a position network, a size network, a framing network, and a recognition network.

The two networks in the part designated by reference numeral 1o in FIG.
The position and size of one pattern present in the input image are each extracted. The two networks are used in parallel, with the position net having only the function of extracting the position of the pattern, and the magnitude net having only the function of extracting the size of the pattern.

The pattern existing in the input image is framed based on the output of the obtained position net and size net.

Next, in order to further improve the accuracy of the frame generated based on the outputs of the position net and the size net, in the part indicated by reference numeral 12, by repeatedly using the framing net several times, it is possible to detect the presence of a pattern. Perform the process of narrowing down the range little by little. Thereafter, the image within the frame corrected based on the output of the framing net is input to the recognition net.

The part indicated by reference numeral 14 recognizes what the pattern is, and the result of judgment based on the output of the recognition net becomes the final answer.

At this time, while changing the frame using the framing net,
It is desirable to view the output of the recognition net several times.

The architecture used here is characterized by separating and simplifying the functions of the four networks. As for the structure of each network, as shown in FIGS. 1, 2, 3, and 5, each network is based on a three-layer BP model. The functions and characteristics of each network will be explained below.

(a) Position net When image information in two-dimensional space is input, one pattern existing in it is extracted and its position coordinates (x,
The network that outputs y) is a position net. Here, the "position" output by the position net is different from the "center of gravity" of the entire input image, which is generally used in pattern recognition. For example, if noise that is largely biased toward the upper left is superimposed, the coordinates of the center of gravity will be far more shifted toward the upper left than the actual pattern position. However, if proper generalization is performed, it is possible to approach the position of the presented pattern even if there is noise.

(b) Size net When image information in two-dimensional space is input, one pattern existing there is extracted and the length d of one side of a square that is large enough to surround it is output. The network is a size net. Based on the output of this size net, a square frame is fitted to the input cover pattern. At this time, we gave a square frame for the vertically long and horizontally long patterns, but we also confirmed that if necessary, the output of the size net could be the length of the vertical and horizontal sides of the rectangle (2 outputs). ing.

(C) Framing Net A framing net is a network that has the function of correcting the frame generated based on the outputs of the position net and the size net, and narrowing down the range in which the pattern exists. It takes the image information inside the frame as input, and outputs the results of determining whether the pattern existing within the frame hits the frame in each of the up, down, left, and right directions. If the output of the unit corresponding to each direction is close to 1, the pattern is hitting the frame, so the frame is moved inward; otherwise, the frame is moved outward.

(d) Recognition Net A recognition net is a network that categorizes the patterns that exist within the fitted frame. If the image information within the frame created by section 10.12 in Fig. 4 is input, the category of the pattern existing there will be output.

By the way, in order to demonstrate the effectiveness of the pattern recognition architecture based on position net and size net,
We conducted a computer simulation to recognize handwritten digits from 5" to 5", and the results are described below.

The networks used in this simulation all have a three-layer structure, and in the intermediate layer and output layer, the output function of each unit is defined as a sigmoid function f (x) −1
/ (1+exp(-x)1) The number of units in each layer was selected as shown in Table 1 based on the results of evaluating the film-forming ability.

Input layer Middle layer Output layer position net 36402 Size net 36401 Framing net 36404 Recognition net 64505 Table 1 Here, in the position net and size net, the input image is
Pixel x 6 pixels horizontally, in the frame drawing net, the image inside the frame is 6 pixels x 6 pixels, and in the recognition net, the image inside the frame is 8 pixels vertically x 8 pixels horizontally.
The pixels are weighted and quantized, and then used as input to each network.

With the above settings, a system was configured using each network after learning. After extracting the position and size, the number of corrections using the framing net was 10 times. After that, framing and recognition are repeated alternately, and the recognition net becomes 3
If the same output was produced twice in a row, that was used as the answer and one recognition task was completed.

FIG. 6 is an example of simulation results. Section 6(a)
The figure shows the input to the system, and FIG. 6(b) shows the output for that input, that is, the output of the recognition net.

Looking at the results, it can be seen that the transformation is invariant to position, size, deformation, and noise.

Additionally, it can be seen that it is able to handle items with blocky noises, which are difficult to use with conventional pattern recognition methods, and items written outside the specified area.

"Effects" According to the present invention, since a position net and a size net are provided to respectively detect the approximate position and size of a pattern having an arbitrary shape and size existing on a two-dimensional screen, noise can be reduced. It is possible to perform extremely robust and accurate pattern recognition against overlapping, positional, and size deviations.

[Brief explanation of drawings]

FIG. 1 is a diagram explaining the function and configuration of the position net of the present invention, FIG. 2 is a diagram explaining the function and configuration of the size net of the present invention, and FIG. 3 is a diagram explaining the function and configuration of the framing net of the present invention. FIG. 4 is a diagram explaining the configuration of the pattern recognition architecture of the present invention. FIG. It is a figure which shows the simulation result of pattern recognition using a net and a frame net. 10... Processing using position net and size net, 12
...Processing by framing net, 14...Processing by recognition net

Claims (4)

[Claims] (1) In a BP model that defines connections between neurons and learning rules, a neuron unit of an input layer that inputs an input two-dimensional image divided into a grid as an intensity signal of each grid, and a neuron unit that inputs an input two-dimensional image divided into a grid as an intensity signal of each grid, an output layer neuron unit that outputs the position of the target pattern, and an intermediate layer neuron unit provided between the input layer neuron unit and the output layer neuron unit, and A neural network for representing the position of a target object, characterized in that the neural network detects the position of the pattern having an arbitrary shape and size that exists in the target object. (2) In a BP model that defines connections between neurons and learning rules, a neuron unit of an input layer that inputs an input two-dimensional image divided into a grid as an intensity signal of each grid, and a neuron unit that inputs an input two-dimensional image divided into a grid as an intensity signal of each grid, an output layer neuron unit that outputs the size of the target pattern; and an intermediate layer neuron unit provided between the input layer neuron unit and the output layer neuron unit, and the input two-dimensional image A neural network for representing the size of a target object, characterized in that it detects the size of the pattern having an arbitrary shape and size that exists on the target object. (3) In a BP model that defines connections between neurons and learning rules, a first input layer neuron unit inputs an input two-dimensional image divided into a grid as an intensity signal for each grid; a neuron unit in a first output layer that outputs the size of a target pattern of a dimensional image, a neuron unit in the first input layer, and a neuron unit in the first input layer;
a position net for detecting the size of the pattern having an arbitrary shape and size existing on the input two-dimensional image; a second input layer neuron unit that inputs an input two-dimensional image divided into a grid as an intensity signal for each grid; and a second output layer neuron unit that outputs the size of a target pattern of the input two-dimensional image. a neuron unit, and a second intermediate layer neuron unit provided between the first input layer neuron unit and the second output layer neuron unit,
and a size net for detecting the size of the pattern having an arbitrary shape and size existing on the input two-dimensional image. (4) In a BP model that defines connections between neurons and learning rules, a first input layer neuron unit inputs an input two-dimensional image divided into a grid as an intensity signal for each grid; a neuron unit in a first output layer that outputs the size of a target pattern of a dimensional image, a neuron unit in the first input layer, and a neuron unit in the first input layer;
a position net for detecting the size of the pattern having an arbitrary shape and size existing on the input two-dimensional image; a second input layer neuron unit that inputs an input two-dimensional image divided into a grid as an intensity signal for each grid; and a second output layer neuron unit that outputs the size of a target pattern of the input two-dimensional image. a neuron unit, and a second intermediate layer neuron unit provided between the second input layer neuron unit and the second output layer neuron unit,
A size net detects the size of the pattern having an arbitrary shape and size existing on the input two-dimensional image, and the input two-dimensional image divided into grids is input as an intensity signal of each grid. Third output that sets the teacher value to 0 when moving the frame inward and 1 when moving it outward for the neuron unit in the third input layer and each neuron unit corresponding to the four directions (up, down, left, right). and a third intermediate layer neuron unit provided between the third input layer neuron unit and the third output layer neuron unit, and based on the outputs of the position net and magnitude net. 1. A neural network for representing the position and size of a target object, characterized in that the neural network is comprised of a framing net that can narrow down the range of existence of the pattern by correcting the frame generated by the pattern.