Image analysis – Pattern recognition – Feature extraction
Reexamination Certificate
1998-08-05
2001-08-07
Johns, Andrew W. (Department: 2621)
Image analysis
Pattern recognition
Feature extraction
Reexamination Certificate
active
06272244
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates in general to methods for extracting and recognizing a pattern in an image, and relates in particular to a method for extracting a particular pattern with an image, and identifying a pattern and its location concurrently, to enable automatic recognition of characters on a license plate for example, and relates also to a method for judging an abnormality in an image.
This application is based on patent application Nos. Hei 9-212104 and Hei 10-71094 filed in Japan, the contents of which are incorporated herein by reference.
2. Description of the Related Art
Methods for extracting certain features of an image, such as particular patterns formed by characters, from an image view (shortened to image hereinbelow) using information processing apparatus are reported in “Extraction of letters from shaded image”, Journal of Information Processing Society, CV Seminar (Reference A) for example.
Also, conventional pattern recognition methods, especially recognition of license plates, have been performed by a process of preliminary binary conversion of the image, followed by pattern matching. This method is successful if the image is clear and has sufficient resolution, then the characters can be independently binarized so that the results can be analyzed. However, recognition is sometimes difficult or impossible when the image has insufficient resolution or the object is unclear or the images are partially touching, shielded or smeared.
Other methods for pattern matching without binarization are based on correlation factor. This method is able to manage partial image touching and shielding problems, but it is inadequate for dealing with identification problems caused by illumination.
Regarding emergency detection techniques, monitoring cameras are often based on image abnormality detection derived from the above-noted pattern extraction and recognition technologies, therefore, any change in an image can constitute a criterion for abnormality. Judgment that an image contains a change is generally arrived at by computing the difference between a current image and the reference image which would normally be viewed by the camera.
However, although methods which are not adversely affected by shadows are known, as reported in reference A, existing methodologies suffer from a common problem that accurate pattern extraction cannot be carried out in many cases, including the cases when there are noise or shadows in an image or when the pattern to be extracted is locally shielded or when the object is not a character.
With respect to currently available image abnormality detection technologies, when there is obstructive element in the image, shadow of a building for example, these methods will recognize the shadow as an image abnormality thereby generating excessive false warnings.
Also, the extent of abnormality is an important criterion for noting an emergency. However, if a genuine image abnormality exists but it is affected by pseudo effects of building shadow and others, apparent abnormality region in an image and the actual abnormality region do not match and accurate area of the actual image abnormality cannot be established, and the process is unable to reach a correct conclusion.
SUMMARY OF THE INVENTION
It is an object of the present invention to provide a methodology that will enable to reliably extract a target pattern even when there is noise or shadow in an input image or a target pattern in an image is partially shielded. The image recognition technology based on the present methodology enables to reduce the restrictions placed on image recording conditions such as illumination requirements as well as to reduce obstructive effects produced by partially shielded image.
Another object is to provide a recording medium storing the pattern extraction programs to be used in conjunction with information processing apparatus.
Still another object is to provide an abnormality detection method that enables to distinguish pseudo-change in an image, caused by such interference effects as shadow of a building, from a genuine change in the image.
The present method for extracting a pre-selected pattern from an image according to programs in computer means is comprised by the steps of: storing reference patterns for extraction; storing input images; applying filter means to reference patterns and input images; integrating a filtered result of the reference pattern and a filtered result of the input image to generate an integrated result; and extracting a reference pattern according to the integrated result and specifying a location of the reference pattern in the input image.
Filter means are steps to provide a preliminary processing to the image and applying one of either a convolution or first differential processing.
The step of integrating includes a convolution of an inner integral of or an outer integral of or a difference of a reference pattern filter output and an input image filter output.
Filter means also includes a step of weighting of each coordinate point.
In more specific terms, the present method chooses appropriate pre-selected mathematical filters to identify local features in both target pattern and reference patterns and to integrate the filtered results in a macroscopic way, which might be likened to a balloting process.
A local feature is a feature of an image which can be computed using microscopic data in a given local region of a reference pattern.
A macroscopic feature is a feature of an image which can be computed using local features over an entire defined region of a reference pattern.
The present method for recognizing a target pattern, in stored input images, which is similar to a plurality of pre-stored reference patterns and associated pattern locations, is comprised by the steps of: cropping of target regions for determining interim target regions; computing local similarities between the target pattern and the reference patterns in a preliminary processing to obtain intermediate data; computing a degree of resemblance by comparing directional differentiated data of the target pattern and the reference patterns; extracting local patterns according to results of the degree of resemblance to enable selecting similar reference patterns; restoring extracted regions so as to match with an original input image by verifying extracted pattern locations with reference pattern locations; deciding on a matching reference pattern for each extracted region; and outputting recognition results and associated reference pattern locations according to results of comparison.
The target pattern in stored input images include numbers on license plates; and the reference patterns are characters including numerals, dots, kanji, hiragana and alphabet letters in case of Japan.
The present invention includes a computer-readable recording medium for executing a programs to recognize a target pattern in stored input images, comprised by the steps of: cropping of target regions for determining interim target regions; computing local similarities between the target pattern and the reference patterns in a preliminary processing to obtain intermediate data; computing a degree of resemblance by comparing directional differentiated data of the target pattern and the reference patterns; extracting local patterns according to results of the degree of resemblance to enable selecting similar reference patterns; restoring extracted regions so as to match with an original input image by verifying extracted pattern locations with reference pattern locations; deciding on a matching reference pattern for each extracted region; and outputting recognition results and associated reference pattern locations according to results of comparison.
Accordingly, the present method of extracting and identifying a pattern within an input image is based on performing local computational processing first to obtain intermediate data, and a degree of resemblance is calculated using directional derivatives of the intermediate data, and superi
Ohtsuka Sakuichi
Shio Akio
Takahashi Yasuko
Azarian Seyed
Johns Andrew W.
Nippon Telegraph and Telephone Corporation
Pennie & Edmonds LLP
LandOfFree
Methods for extraction and recognition of pattern in an... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Methods for extraction and recognition of pattern in an..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Methods for extraction and recognition of pattern in an... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2515141