CN106097248B - High-resolution image knowledge prior-based compressed sensing method and mixed vision system thereof - Google Patents

Abstract

The invention belongs to the technical field of machine vision, and in particular relates to a compressed sensing method based on high-resolution image knowledge prior and a hybrid vision system thereof. The invention includes: setting the middle dimension of the high-dimensional image and the low-dimensional image in the hybrid vision system; calculating the high-dimensional image dimensionality reduction observation matrix and reducing the dimensionality of the high-dimensional image; dimension; use the SIFT matching algorithm to match the obtained two images of the same dimension and calculate the matching rate; repeat the above steps until the optimal middle dimension that can obtain the highest matching rate is found. Using the means of image processing, aiming at the characteristics of the imaging scale deviation of the target in different visual primitives of the mixed vision system, the compressed sensing technology is used to reconstruct the panoramic nonlinear compressed image and the perspective image is reduced based on the down-sampling measure. dimension.

Description

A Compressed Sensing Method Based on High-Resolution Image Knowledge Prior and Its Hybrid View sensory system

technical field

The invention belongs to the technical field of machine vision, and in particular relates to a compressed sensing method based on high-resolution image knowledge prior and a hybrid vision system thereof.

Background technique

Due to its characteristics of non-contact perception, rich information acquisition and strong anti-interference ability, vision technology has a wide range of applications in the fields of environment understanding, target detection and positioning. Conventional binocular vision and binocular panoramic vision have the disadvantages of narrow field of view and limited distance when performing target positioning. It has significant advantages in the field of target tracking and positioning applications within a large field of view. However, due to the different imaging principles of visual primitives and different target imaging scales in this hybrid system, the classic binocular vision positioning algorithm fails. Effectively solving the problem of equal-dimensional reconstruction of images between images of different scales in a multi-eye stereo vision system composed of heterogeneous imaging vision units will be the most basic and important technology for realizing three-dimensional information perception of the environment based on machine vision. Big promotion.

According to the classic Shannon sampling theorem, in order to restore the analog signal without distortion, the sampling frequency should not be less than twice the Nyquist frequency (ie, the highest frequency in the analog signal spectrum). However, except for the assumption that the signal is of limited bandwidth, no other prior information is used. There is a large degree of redundancy in the collected data. The Compressed Sensing method (Compressed Sensing CS) makes full use of the prior information that most signals can be sparsely represented on a predicted basis, and provides a new idea for the reconstruction of dimensions.

Contents of the invention

The purpose of the present invention is to provide a compressed sensing method based on high-resolution image knowledge prior using machine vision means.

The object of the present invention is also to provide a hybrid vision system based on the compressed sensing method of high-resolution image knowledge prior.

The purpose of the present invention is achieved like this:

A compressed sensing method based on high-resolution image knowledge prior, comprising the following steps:

1) Set the intermediate dimension of the high-dimensional image and the low-dimensional image in the hybrid vision system;

2) Calculate the high-dimensional image dimension reduction observation matrix and reduce the dimension of the high-dimensional image, calculate the low-dimensional image dimension enhancement observation matrix and increase the dimension of the low-dimensional image;

3) use the SIFT matching algorithm to match the obtained two pictures of the same dimension and calculate the matching rate;

4) Repeat the above steps until the optimal intermediate dimension that can obtain the highest matching rate is found.

A hybrid vision system based on the compressed sensing method of high-resolution image knowledge prior, the upper part is a panoramic camera, and the lower part is a perspective camera, wherein the panoramic camera includes a hyperboloid omnidirectional imaging mirror 1, through the panoramic The camera annular light-transmitting glass support tube 2 is supported on the panoramic camera support base 3, and on the panoramic camera support base 3 in the panoramic camera annular light-transmissive glass support tube 2 below the omnidirectional imaging reflector 1, 1394 cameras 4 are arranged vertically upward , the 1394 camera 5 as a perspective camera is vertically linked to the rotatable link rod 6 and vertically linked on the panoramic camera support base 3, and the perspective camera annular light-transmitting glass support cylinder 7 supports the panoramic camera and is placed on the perspective camera support base 8.

The beneficial effects of the present invention are:

Using the means of image processing, aiming at the characteristics of the imaging scale deviation of the target in different visual primitives of the hybrid vision system, the compressed sensing technology is used to reconstruct the panoramic nonlinear compressed image and the perspective image is reduced based on the down-sampling measure. dimension.

Description of drawings

FIG. 1 is a basic block diagram of a compressed sensing technology based on high-resolution image knowledge prior in the present invention.

Fig. 2 is a schematic diagram of the hybrid vision device of the present invention.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings.

The invention discloses a compressed sensing technology based on high-resolution image knowledge prior. The main steps are: setting the intermediate dimension of high-dimensional image and low-dimensional image in mixed vision, calculating the high-dimensional image dimensionality reduction observation matrix and performing high-dimensional Image dimensionality reduction, calculate the observation matrix of low-dimensional image enhancement and increase the dimensionality of the low-dimensional image, use the SIFT matching algorithm to match the obtained two images of the same dimension and calculate the matching rate, repeat the above steps until the highest matching rate can be obtained The optimal middle dimension of . The invention adopts the scheme of machine vision and image processing, and utilizes the theory of compressed sensing to realize the reconstruction and mapping of low-dimensional panoramic target feature images to high-dimensional images, and at the same time perform down-resolution sampling on conventional large-scale images, combined with the matching accuracy test results , to find the best matching point between panoramic feature scale up and perspective feature scale down, realize the precise matching algorithm of the target features at different scales, and improve the positioning accuracy of 3D information perception in mixed visual environments.

The technical solution adopted by the present invention to achieve the purpose of the invention is: there are large differences in the imaging scales of the same target of interest in the heterogeneous dual-scale hybrid vision system, mainly due to the spatial light domain compression effect in the imaging process of the panoramic vision through the hyperboloid mirror, The corresponding relationship between the spatial field of view and the imaging area shows significant nonlinear characteristics. On the contrary, conventional vision system imaging can perform fine imaging of the target through optical zoom operation. Therefore, how to achieve accurate matching of features in images of different scales is an important step in positioning . Since the panoramic system has "complete" and "continuous" characteristics for environmental imaging, and small-scale CCD digital discrete imaging is "discrete compressed sampling" of spatial imaging information, using the theory of compressed sensing to realize low-dimensional panoramic target feature images to high-level At the same time, the conventional large-scale image is down-sampled, combined with the matching accuracy test results, the best matching point is found for the scale-up of the panoramic feature and the dimensionality reduction of the perspective feature, so as to realize the target features at different scales. Exact match algorithm.

A kind of compressed sensing technique based on high-resolution image knowledge prior, it is characterized in that: comprise following main steps:

1) Estimate and set the intermediate dimensions of high-dimensional images and low-dimensional images in hybrid vision based on prior knowledge.

2) Calculate the high-dimensional image dimensionality reduction observation matrix and reduce the dimensionality of the high-dimensional image, calculate the low-dimensional image dimension-up observation matrix and increase the dimensionality of the low-dimensional image.

3) Use the SIFT matching algorithm to match the obtained two images of the same dimension and calculate the matching rate.

4) Repeat the above steps until the optimal intermediate dimension that can obtain the highest matching rate is found.

Claims (2)

1. A compressed sensing method based on high-resolution image knowledge prior is characterized by comprising the following steps:

1) Setting the middle dimension of a high-dimensional image and a low-dimensional image in a hybrid vision system;

2) calculating a high-dimensional image dimension reduction observation matrix, reducing the dimension of the high-dimensional image, calculating a low-dimensional image dimension increasing observation matrix, and increasing the dimension of the low-dimensional image;

3) Matching the two obtained pictures with the same dimensionality by using an SIFT matching algorithm and calculating a matching rate;

4) Repeating the steps until the optimal middle dimensionality which can obtain the highest matching rate is found.

2. the hybrid vision system applying the high-resolution image knowledge prior compressed sensing method according to claim 1, wherein: the panoramic camera comprises a panoramic camera body, the upper half part is a perspective camera, the panoramic camera body comprises a hyperboloid all-dimensional imaging reflector (1), the hyperboloid all-dimensional imaging reflector is supported on a panoramic camera support base (3) through a panoramic camera annular transparent glass supporting cylinder (2), a 1394 camera (4) is vertically and upwards arranged on the panoramic camera support base (3) in the panoramic camera annular transparent glass supporting cylinder (2) below the all-dimensional imaging reflector (1), a 1394 camera (5) serving as the perspective camera is vertically connected to a rotatable link rod (6) and is vertically connected to the panoramic camera support base (3), the panoramic camera is supported by the perspective camera annular transparent glass supporting cylinder (7) and is placed on a perspective camera support base (8).