JPH01180069A - Device for extracting image feature - Google Patents

Abstract

PURPOSE:To recognize whether an image is a dot-picture or not and to cause an image processing to be easy by obtaining Hadamard's transform to binarizing data, which are binarized in a binarizing means, in an Hadamard's transform means and executing feature extraction with the Hadamard's transform in a feature extracting means. CONSTITUTION:Picture data 151, which are sampled and caused to be a digital signal by a scanner, etc., are binarized by an image binarizer 101 and binarizing data 152 are temporarily accumulated in a binarizing data buffer 102 so that a prescribed main scanning number can be simultaneously read. These accumulated picture data 153 are inputted to an Hadamard transformer 103 and Hadamard's transform 154 is generated to the binarizing picture of respective picture elements in a sub-scanning direction and a main scanning direction. This Hadamard's transform 154 is inputted to a feature extracting part 104 and binarization and logic operation are executed to one part of the Hadamard's transform 154. Then, a feature 155 is outputted.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field] The present invention relates to an image feature extraction device, and more particularly to an image feature extraction device that obtains a Hadamard transform of an image and recognizes the properties of the image based on the values of the transform.

[Conventional technology]

Conventionally, this type of device has handled image data as multivalued data, in most cases 8 bits/pixel. Therefore, since the scale of the device would be large, the device was not created using dedicated hardware, but was realized using computer software. That is, there are disadvantages in that the processing time is long and a huge memory is required to store images and converted images that cannot be processed in real time.

[Purpose of the invention]

The present invention aims to provide an image feature extraction device that eliminates the drawbacks of the above-mentioned conventional examples, enables real-time processing, can perform processing without requiring a huge memory, and has a simple circuit and is easy to implement in hardware. There is.

〔Example〕

FIG. 1 shows an embodiment of the present invention, in which 101 is an image binarizer;
02 is a binary image data buffer, 103 is a Hadamard transformer, and 104 is a feature extractor. This will be explained below with reference to FIG. Image data 151 sampled by a scanner (not shown) and converted into a digital signal is binarized by an image binarizer 101.

Binarized data 152 is stored in binary image data buffer 1.
02, data for m main scans are temporarily stored so that they can be read simultaneously. m pieces (m bits) of image data 153 are input to the Hadamard transformer 103, and a Hadamard transform 154 is generated for a binary image of m pixels in the sub-scanning direction and n pixels in the main scanning direction. Hadamard transformation 154
is input to the feature extractor 104 and the Hadamard transform 154
Binarization and logical operations are performed on a portion of , and a feature 155 is output. Each part will be explained in detail below.

FIG. 2 is a block diagram of an embodiment of an image binarizer.

201 is a smoother that blunts the signal, 202 is a signal delay device, and 203 is a comparator. This will be explained according to FIG.

Signal 15 digitized by a scanner (not shown)
1 is input to the signal smoother 2011 and the signal delay device 202. Inside the signal smoother, although not detailed here,
For example, using a 3x3 mask as shown in Figure 8,
Performs dimensional convolution to dull the signal. On the other hand, the signal delayer 202 delays the signal by the time required for smoothing, so that the smoothed signal 252 and the delayed original signal 253 are at the same position with respect to the coordinates of the image. Two signals 252 and 253 are generated as described above,
It is input to comparator 203. The comparator outputs 1 as the binary image signal 152 when the signal 253 is greater than the signal 252, and 0/ otherwise. Here, it is defined that when the value of the image is convex or increasing, it becomes l.

The binary image signal 152 is buffered by the binary image buffer 102 so that m pixels in the sub-scanning direction can be referenced simultaneously. n pixels in the main scanning direction, m pixels in the sub-scanning direction (mXn)
Binary partial image data 1 of partial image block consisting of pixels
53 is input to the Hadamard transformer 103, and a Hadamard transform 154 is generated. FIG. 3 is an explanatory diagram of the process described above, in which 301 is the original image, 302 is an enlarged display of the original image,
303 shows the Hadamard transformation. 30
4.306 each indicate one mXn partial image block, and 305.degree. 307 indicate Hadamard transforms corresponding to 304 and 306, respectively. 351 is the main scanning direction of the image scan, and 352 is the sub-scanning direction of the image scan. This will be explained according to FIG. One m x n partial image block 306 of the image 301 is input to a Hadamard transformer, not shown in FIG.
Get 7. The m×n subimage block 304 is then processed to obtain the Hadamard transform 305. Thereafter, the rask scan is performed in units of partial image blocks in this manner, and the entire image is processed. The processing of one m x n partial image block will be described below. FIG. 4 shows a block diagram of an embodiment of the Hadamard transformer 103. In Figure 4,
401 is a Hadamard coefficient generator, 402 is a bidirectional counter (which can be increased or decreased), 403 is a parallel/serial converter, 40
4 is a serial/parallel converter. One partial image block data 153 from the binary image buffer memory 102 is
The parallel to serial converter 4 is activated by the l block processing start signal 458.
It is latched to 03. Then, or simultaneously, the counter clear signal 455 causes the parallel-to-serial converter 403 . Hadamard coefficient generator 4o1. Bidirectional counter 402 is initialized. One pixel (1 bit) of data 451 is output from the parallel/serial converter 403 in one cycle of the pixel clock 454.
, 1-bit coefficient data 456 is read out from Hadamard coefficient generator 401 and input to bidirectional counter 402 . The bidirectional counter 402 operates as shown in Table 1 below, and one Hadamard transform is generated in mXn cycles of the pixel clock 454. Here, the Hadamard coefficient will be explained. The Hadamard coefficient used in this embodiment has two special characteristics. One is 1. 1.0 instead of -1
The other is to generate a direct product of two Hadamard matrices to facilitate the two-dimensional Hadamard transformation. FIG. 9 shows an example of an 8th order Hadamard matrix. The Hadamard coefficient used in the present invention is written as HC- (H" (+, 1) is the l, J element of the m-dimensional Hadamard matrix as can be inferred from Figure 9) HC-a, Lk, l) = H”(1,i)■H”(k,
g, ■ are EXCLUSIVE OR.

It can be written as Hadamard coefficient generator 401 in FIG.
Signals 452 and 453 inputted to the above correspond to i and l in the above-mentioned equations, respectively, and these are parameters for generating Hadamard coefficients for sequence -β in the main scanning direction and sequence -1 in the sub-scanning direction. The value corresponding to Lk in the above equation is degenerated by being calculated with the corresponding two-dimensional image using the pixel clock 454. It has been explained above that the bidirectional counter 402 generates the Hadamard transform 457 of the sequence indicated by the signals 452 and 453. Hadamard transform 457 is a series
It is input to a parallel converter 404, and when a two-dimensional Hadamard transform is determined, it is output as a Hadamard transform 154. Although the generation device for the parameters 452 and 453 was not made clear in the above description, it is clear that they can be configured by counters. The Hadamard transform was found for the above m x n partial image. Figure 5 schematically shows the Hadamard transformation, where hi (i, j)
indicates elements of Hadamard transform whose sequences in the sub-scanning and main-scanning directions are i and j, respectively, and hatched elements indicate that the absolute value of the value is large.

Hereinafter, m=8. Feature extraction will be explained assuming n=32.

FIG. 5 also shows an example of Hadamard transform for character line images. On the other hand, FIG. 6 shows an example of Hadamard transform for a 100 lines/inch, 45 degree halftone image. Figure 5,
In both FIG. 6, when the absolute value of the Hadamard transform element is greater than 20, it is shown by hatching, and when it is less than 20, it is shown in white. A clear difference is observed.

In this embodiment, these differences are summarized as follows.

1.7 When the absolute value of the element of Damard transform is greater than 40, the map becomes 0/ when it is less than 1.40. If it is not 0/, the corresponding partial image block is recognized as a halftone dot.

2. When the absolute value of the elements of the Hadamard transform is greater than 20, create a map that becomes 0/ when it is less than or equal to 1.20. If not, the corresponding partial image block is recognized as a halftone dot.

Examples of masks A and B are shown in FIG.

Here, 701 is mask A, and 702 is mask B.

FIG. 10 shows an embodiment of the feature extraction described above.

In FIG. 10, 1001 is a parallel-serial converter, 100
2 is a mask A generator, 1003 is a mask B generator, 10
04゜10051i comparator, 1oo6, 1oo7haAN
A D circuit, 1008 is an OR circuit, and 1009 is a D flip-flop.

l block start signal 458 causes parallel to serial converter 100
The Hadamard transform is latched at 1. At the same time, mass
The reset generators 1002, 1003 are initialized and the flip-flop 1009 is cleared. One element of the Hadamard transform is read out in one cycle of the pixel clock 454,
Signal 1051 and comparator 1004. 10051:
m inputs, each of which is binarized with a threshold value of 40.20.

At this time, the absolute value of signal 1051 is used for comparison.

Binarized signals 1054 and 1055 are elements 1052 and ? of mask A, respectively. 7. (7) Element 1053 is ANDed and stored in flip-flop 1009 through 0RI008. At this time, if there is any element among the elements mXn that is determined to be a halftone dot, 1009 remains set to 1 and is not changed to O/. As explained above, it is possible to sequentially determine for a partial image block of m x r (pixels) whether or not the image block has the characteristic of a halftone dot.

Although all of the embodiments described above are constructed using hardware circuits, it is possible to replace all or part thereof with a computer and software.
In particular, if the feature extractor 104 is configured so that mask A, mask B, and threshold values can be changed, special patterns can be extracted and the number of lines is limited.

It is also possible to extract only halftone images with angles.

〔effect〕

As explained above, by calculating the Hadamard transform of a partial small area of binarized image data and extracting the features, it is possible to recognize whether or not the image is a halftone image, which facilitates image processing. It has the effect of making

[Brief explanation of the drawing]

Fig. 1 is an overall block diagram of an embodiment of the present invention, Fig. 2 is an embodiment of an image binarizer, Fig. 3 is a diagram explaining Hadamard transform processing of this embodiment, and Fig. 4 is a Hadamard transform. 5 and 6 are block diagrams of the embodiment of the Hadamard transform.
Figure 7 shows a mask used for feature extraction; Figure 8 shows an example of a convolution kernel used in an image binarizer; Figure 9 shows an example of a Hadamard transformation matrix. , FIG. 10 is a circuit diagram of an embodiment of a feature extractor.

Claims (1)

[Claims] Binarization means for binarizing image data; Hadamard transformation means for performing Hadamard transformation on the binarized data;
An image feature extraction device comprising feature extraction means for extracting features by Hadamard transformation.