Overview

Static image recognition is a central issue in Deep Learning and others. However, when we apply deep learning to an image which includes images belonging multiple categories and overlapped each other, and also their shapes and textures are non-linearly deformated, it is difficult to recognize each of them.
 
 It is said that deep earning is currently difficult to introduce "hierarchy". This is the main reason why deep learning has the weakness when facing the problem mentioned above.
 
 Hierachy should be introduced for solving the segmentation of image. However, the problem of segmentation and recognition are strongly coupled each other. That is, "chicken and egg" problem exists. Recognition becomes easier if segmentation is possible, and vice versa. It is necessary that we develop an algorithm which can solve "chicken and egg" problem that is the strong coupling problem of segmentation adn recognition

  
Two-dimensional continuous DP (2DCDP) is extension of a one-dimensional continuous DP to two dimensions. One-dimensional continuous DP (CDP) was proposed by Oka in 1978, which realizes segmentation-free recognition of a one-dimensional pattern such as time series.

For two-dimensional continuous DP, it is applied independently for each input image using each individual reference image with a single category, and segmentation-free recognition becomes possible. This is because each individual reference pattern constitutes linear combination of hierarchy, while multiple category images consitute a single hierarchy in Deep Learning. Also, since the two-dimensional continuous DP absorbs nonlinear deformation including the enlargement / reduction of size of the individual target image, the reference pattern ("learning data") is enough to take one.
 
  Two-dimensional continuous DP is useful for recognizing overlapped amd multiple image patterns in a target image by taking such a linear combination of hierarchical structure, that is, 2DCDP applies independently using a single reference image.

One dimensional DP is naturally expanded to two-dimensional image pattern by a series of joint works of many researchers, Dr. T.Nishimura (now AIST), Mr. Iwasa(now Seiko-Epson Inc.), Dr. Y.Yaguch (U of Aizu) with me. Finally, the algorithm becomes a quite sophistcated and almost final version of 2DCDP.

On the other hand, 2DCDP requires only one reference (learning) pattern, while DL requires a large amount of data for learning.

The one-dimensional Continous Dynamic Programming was published is the follwing
paper. However, the first paper of CDP was publised in 1978 in Japanese.


[1] "Spotting Method for Classification of Real World Data": Ryuichi Oka, The Computer Journal, Vol.41, No.8, pp.559-565 (1998)

Laboratories and Groups

Research Category

Areas of Activity
Image Processing
Related Keywords