JPS59214987A - Extracting device of picture feature - Google Patents

Abstract

PURPOSE:To make feature extraction easy to eliminate the redundancy of data by dividing a recognition object picture to patterns similar to this picture and extracting features from these similar pictures. CONSTITUTION:A dividing part 3 is controlled by a control part 5 and does not divide first the recognition object picture, which is outputted from a normalizing part 2-3, to output this picture as it is, and hereafter, the dividing part 3 divides this recognition object picture to a prescribed number of partial pictures similar to the recognition object picture as required to output them. A feature extracting part 4-2 extracts first features from the recognition object picture which is not divided, and next, the extracting part 4-2 extracts features from divided partial pictures successively as required. Thus, rough features are extracted first from the recognition object picture which is not divided, and next, local features are extracted from partial pictures successively as required. The feature extracting part 4-2 uses the same method for the recognition object picture which is not divided as well as partial pictures.

Description

Detailed Description of the Invention (5) Technical Field of the Invention The present invention relates to an image recognition device.
′: Regarding.

(Bl Technological Background 1N!Il In the image journal industry, the method is generally to extract the % features of the image to be recognized and compare the features with standard features previously calculated from known images and stored in a dictionary. is used, but because the target is two-dimensional information, the amount of data handled is large, and a large-capacity storage device is required for the dictionary.
Also, it takes time to extract features, match them, etc.
Attempting to reduce these is an important technical issue in recognition devices. (C) Prior art and problems In the image recognition process, conventionally, one type of feature is extracted and matched. However, in the former case, in order to obtain stable recognition results, it is necessary to extract features in detail, and therefore feature extraction is complicated. It had the disadvantage of becoming -1: In the latter case, even in cases where recognition is relatively easy, multiple SD features are provided, resulting in data redundancy, which is a drawback.

1) Purpose of the Invention The purpose of the present invention is to eliminate the problems in the conventional example, that is, to simplify target feature extraction and eliminate data redundancy.

(G) Structure of the Invention The image feature extraction device according to the present invention includes means for dividing a recognition target image into a similar number of partial images to the recognition target image;
means for extracting the features of any one of the recognition target image and the similar number of partial images;
By dividing the recognition target image into figures similar to the recognition target image and extracting features from the similar images, feature extraction is facilitated, and if necessary, local features of the recognition target image are extracted from the divided images. By extracting the data, redundancy is removed.

EXAMPLES OF THE INVENTION Hereinafter, the essence of the present invention will be specifically explained by referring to illustrated examples.

11 Figure 6 The present invention-embodiment jI shows the configuration, lI′i input section containing the recognition target image) Recommend image A 1;, 512
X512) 5 input parts, 2-1, 2-2, and 2-3, which convert into a dot x 8-bit digital image, constitute a preprocessing part 2, and 2-1 converts into a digital image converted by the input part 1. a noise removal unit 2-z that removes noise from the digital image by performing median filter processing on the digital image;
Edges are detected by performing differential filter processing on the digital image output by the noise removal unit 2-1, and the edges are integrated using known knowledge regarding the shape of the recognition target image to extract the contour line of the recognition target image. A contour extraction unit that
2-3 is 256
256) A normalization unit that normalizes the image into a dot x 8-bit digital image;
a dividing unit 4-1, which outputs the recognition target image outputted by 3 without dividing it at first, and thereafter divides the recognition target image into partial images having titration numbers similar to the recognition target image as necessary; 4-2 constitutes a means for extracting the features of the recognition target image and a plurality of partial images similar to the recognition target image, and 4-1 outputs the recognition target image output by the dividing unit 3 as it is and divides the recognition target image into a partial image. Perform coordinate transformation including rotation, translation, and scale transformation on the image 256 x 256
4-2 is a normalization unit that normalizes and outputs a dot x 8-bit digital image it, 4-2 is a feature extraction unit that extracts percentage features of the recognition target image and partial image output by the normalization unit 4-1, and 5 is a control section that controls each section.

The second N is for the case where the outline of the figure to be recognized is a square.
An example of a partial image that is detected in the dividing unit 3 is shown below, and (a
) is the recognition target image, (b) is a partial image similar to the recognition target image obtained by dividing the recognition target image into four equal parts, and (C) is the recognition target image obtained by dividing the recognition target image into nine equal parts. This is a partial image similar to . It also shows the projection of density onto two sides that are momentarily connected to each image.

The % weight extraction unit 4-2 projects the density onto two sides that are instantaneously connected to each figure, and obtains a projected image represented by two 256-dimensional vectors. Each of these is equally divided into 32 sections Ikl, and the sum is calculated for each valley section. Next, by calculating the difference between the sums for each adjacent section, features expressed by two 31-dimensional vectors are extracted for each figure.

The feature extraction carried out by the % fine extraction unit 4-2 is first carried out on the undivided recognition target image, then as necessary on the partial images of each part obtained by dividing into four parts, and further as necessary. 9 divisions, 16 divisions, mountains,
This is done for each partial image obtained.

In this way, redundancy in the extracted feature data can be removed by first extracting global features using the recognition target image, which is divided into parts, and then sequentially extracting local features using partial images as necessary. In addition, the feature extraction unit 4-
'2 uses the same method for undivided recognition target images and partial images, and detailed features can be extracted from 1-minute autopsy images, which simplifies the feature extraction part and improves feature extraction. The time required can be shortened.

In the embodiment described above, since the image supplied to the feature extraction section 4-2 is normalized, the movement of the feature extraction section 4-2 can be particularly simplified.

(Gl As described in detail, according to the present invention, feature extraction in an image recognition device is performed by f≦1, eliminating redundancy of one take,
By doing this, the capacity of the storage tank and the time required for preparation, extraction, heating, etc. can be reduced.

[Brief explanation of drawings]

Figure 1d is a diagram showing the deterioration of an embodiment of the present invention, and Figure 2 is a diagram showing an example of division of a figure to be recognized. In FIG. 1, 211'' is a J cutting processing section, 3 is a dividing section, 4-1 is a straight arrival unitizing section, and 4-2 is a % extraction section.

Claims (1)

[Claims] (1) A means for dividing a recognition target image into a plurality of partial images similar to the recognition target image, and a means for extracting image characteristics of which of the recognition target image and the plurality of similar partial images. (2+) Claim 1 of the patent characterized in that the means for extracting the percentile of the image comprises means for normalizing the partial image to a predetermined size. The image feature extraction device described.