A kind of road information Remotely sensed acquisition method based on Pulse Coupled Neural Network method
Technical field
The present invention relates to a kind of high resolving power urban road information remote sensing extracting method based on coupled pulse neural network.
Background technology
Along with the fast development of satellite remote sensing and computer technology, the resolution of remote sensing image is more and more higher.How to extract Target scalar quickly and accurately, become the important proposition of current remote sensing image processing.Urban road, as important ground object target, is the important component part of geographic information database.The real-time update of urban road is significant for automobile navigation, traffic administration, city planning and urban study simultaneously.Therefore, carry out urban road high-definition remote sensing extraction research and there is important significance of scientific research and practical value.
Urban road, due to the interference by multiple atural object, is often difficult to accurate extraction.Therefore Chinese scholars has done large quantity research, proposes certain methods.Along with remote sensing image processing method development and the diversification of data can be obtained, the future development that combines to multi-method and multi-data source of road extraction technology in recent years.Multi-method in conjunction with in, the method that Movaghati utilizes particle filter and EKF to combine carries out urban road extraction.The method that Maurya utilizes K mean cluster and mathematical morphology to combine carries out road extraction.Thunder is little very proposes a kind of method for extracting roads combined with shape facility based on the conforming Iamge Segmentation of local gray level.Zhou Shaoguang proposes the high-resolution remote sensing image road segment segment extracting method that a kind of Shape-based interpolation priori and figure cut.Multi-data source in conjunction with in, Herumurti uses take photo by plane data and DSM data of optics to carry out road extraction.Li Yijing merges high resolution image and LiDAR data, and the road achieved under complex scene extracts automatically.Above new method has certain using value, and for road extraction research provides significant reference, but multi-method combines, to relate to technology more, to a certain degree has influence on the integrated of algorithm and implementation efficiency.The data type that the method for multi-data source relates to is more, therefore there is the problem that data acquisition is difficult and cost is higher.
Consider above problem, the feature and the coupled pulse neural network that the present invention is directed to road can ignore gray difference little in the same area and space interruption, thus the advantage of complete road information can be obtained, the method for extracting roads based on Pulse Coupled Neural Network is proposed.The proposition of the method is also for the extraction of river information provides quite significant reference.
Summary of the invention
Comparatively large for road extraction difficulty, existing method for extracting roads fits the problems such as general property is poor, the invention provides the comparatively strong and road information extractive technique flow process that precision is higher of simple, the suitable general property of a kind of method.
Object of the present invention is realized by following technical step:
Step 1) necessary pre-service is carried out to image, namely use principal component analytical method to obtain the image of the first component of original high resolution image, and use adaptive histogram equalization method to carry out image enhancement processing.
Step 2) utilize the method for Pulse Coupled Neural Network to the Image Segmentation Using process after enhancing, tentatively obtain road information.
Step 3) utilize the near red wave band of original image and red wave band to obtain normalized differential vegetation index figure, and use the image of the figure image subtraction normalized differential vegetation index after segmentation, namely get rid of the impact of vegetation in segmentation image.
Step 4) use the methods such as mathematical morphology to carry out aftertreatment, stress release treatment disturbs, and connects fracture road, thus obtains final road.
Further, described step 1) concrete grammar be:
A) based on principal component analytical method, original image is carried out principal component transform, original image is converted into each component image incoherent mutually, and the image getting the first component is as image to be dealt with; B) use adaptive histogram equalization method to carry out image enhancement processing to the image of the first component, increase the otherness of road information and background information.
Further, described step 2) concrete grammar be:
A) relevant input parameter and the iterations of coupled pulse neural network is obtained according to feature of image and experiment; B) according to the Related Formula of coupled pulse neural network, image is processed; C) preliminary road information is obtained according to the iterations determined.
Further, described step 3) concrete grammar be:
A) according to handled different remote sensing images, find the near-infrared band of image and red wave band, obtain the normalized differential vegetation index image of image according to the formula of normalized differential vegetation index; B) binary conversion treatment is carried out to the image of the normalized differential vegetation index obtained, obtain the binary map of normalized differential vegetation index; C) according to circumstances to the process using the image of Pulse Coupled Neural Network method process to carry out the upset of meeting degree, guarantee that reason information remains white on image; D) image of the normalized differential vegetation index after the figure of the Pulse Coupled Neural Network after process and process is carried out making difference to process, eliminate the disturbing effect of vegetation in Pulse Coupled Neural Network image.
Further, described step 5) concrete grammar be:
A) image eliminating vegetation factor is carried out to the process of area opening operation, namely eliminate the noise spot of small size; B) utilize Mathematical Morphology Method to carry out the image after to the process of area opening operation to process, thus reach connection fracture road, fill up the object of road Hole.
Accompanying drawing explanation
Fig. 1 is raw data;
Fig. 2 is the first component after principal component transform;
Fig. 3 is image enhaucament result;
Fig. 4 is Pulse Coupled Neural Network method segmentation result;
Fig. 5 is normalized differential vegetation index figure;
Fig. 6 is for eliminating Vegetation Effect design sketch;
Fig. 7 is road extraction result figure;
Embodiment
Below in conjunction with accompanying drawing, explanation is further elaborated to the present invention's " a kind of road information Remotely sensed acquisition method based on coupled pulse neural net method ".
(1) Image semantic classification
First, the method for principal component analysis (PCA) is utilized original Multi-Band Remote Sensing Images (Fig. 1) to be transformed to another group each component incoherent mutually.Original image is after principal component transform, and the information content forming each component in major component image is different to a great extent, therefore makes part atural object more outstanding on some component, in addition each component is mutually vertical, increase class spacing, reduce class internal diversity, improve nicety of grading.And the first principal component component that principal component transform produces is equivalent to the weighted sum of original each wave band, what comprise contains much information, the interference being subject to noise is minimum, therefore enhancing and the analysis of detail characteristic is conducive to, be applicable to high-pass filtering, linear feature enhancing and the process such as extraction and density slice, significant for Objects extraction.Therefore this invention processes based on first component map (Fig. 2) of principal component transform.Adaptive histogram equalization method is the innovatory algorithm of histogram equalization, and this algorithm, by the local histogram of computed image, then redistributes the contrast that brightness changes image.Therefore, this algorithm more can be applicable to the local contrast of improvement image and can obtain more image detail.Therefore this patent uses adaptive histogram equalization to carry out image enhancement processing to the first component map, obtains the image (Fig. 3) after increasing, with outstanding road edge details, is convenient to next step Iamge Segmentation process.
(2) coupled pulse neural network lane segmentation
Coupled pulse neural network is the third generation artificial neural network that the brain visual cortex synchronizing pulse granting phenomenon of foundation cat proposes.Introduce after image processing field through Johnson, applied all very well in Iamge Segmentation, image co-registration and image thinning etc.Coupled pulse Application of Neural Network is when Iamge Segmentation, the neuron of its similar input has the characteristic of synchronous spike pulse, the subtle change in the discontinuous and amplitude in space of input data can be made up, thus more intactly can retain the area information of image, obtain good image segmentation.Mathematical iterations equation is expressed as (1) ~ (5):
U
ij(n)=F
ij(n)(1+βL
ij(n))(3)
Y
ij(n)=step(U
ij(n)-θ
ij(n))(5)
In formula, ij subscript represents neuronic label, S
ij, F
ij, L
ij, U
ij, θ
ijbe expressed as the outside stimulus of neuron ij, feeding input, link input, excited inside and dynamic threshold.M is connection weight matrix, VF, VL, V
θrepresent amplitude constant, β is link coefficient, α
f, α
l, α
θfor corresponding time attenuation constant, n is iterations, Y
ijfor exporting.
Neuron is made up of three parts: feed-in unit, linkage unit and impulse generating unit.External signal enters from feed-in unit, and wherein feed-in unit is made up of two passages, and one is F passage, and another is L passage.F passage inputs for the feeding of receiving package containing external input signal, sees formula (1).L passage then for receiving from other neuronic link inputs, is shown in formula (2).At linkage unit, first by link input and link multiplication, and do the skew of constant 1, being then multiplied to modulate with feeding input produces neuronic interior movable item, sees formula (3).Impulse generating unit is mainly made up of threshold adjuster, comparer and impulse generator.Threshold value changes with the change exported, when neuron exports a pulse, and neuronic threshold value θ
ijjust improved by feedback, see formula (4).As neuron threshold value θ
ijmore than U
ijtime, pulse generator will be closed, and stops sending pulse.Threshold value start index declines, when it is lower than U
ijtime, pulse generator is opened, and neuron is lighted a fire, and starts to produce pulse, sees formula (5).If the pixel corresponding to neuron of adjacent neuron and a front iteration point fire has similar intensity, then adjacent neuron is easily lighted a fire.Thus, any one neuronic autogenous ignition all can trigger collective's igniting of its adjacent similar neural unit, and the neuron of these igniting forms a neuron colony, corresponding to the zonule in image with similar quality.Therefore, utilize the Pulse Coupled Neural Network method similarity collective characteristic of catching of lighting a fire just can carry out Iamge Segmentation, thus obtain the preliminary segmentation result (see Fig. 4) of road.
(3) normalized differential vegetation index figure making and eliminate vegetation interference
Vegetation is a kind of important atural object, also creates certain impact to the result of road extraction, therefore needs the interference getting rid of vegetation.This patent obtains vegetation distribution with normalized differential vegetation index, and obtains normalized differential vegetation index figure, and in order to keep being consistent with the form of Pulse Coupled Neural Network segmentation result, normalized differential vegetation index figure is converted into binary map (Fig. 5) the most at last.Then by the subtraction of image, the normalized differential vegetation index figure after the image of Pulse Coupled Neural Network process and binaryzation is carried out making difference operation, eliminate the interference of vegetation, the figure (Fig. 6) after the vegetation that is eliminated.
(4) post processing of image
Post processing of image is an important ring of road extraction, because the image after Pulse Coupled Neural Network process makes road information comparatively complete, considers that usable floor area opening operation removes other isolated noises, thus obtains road information more accurately.Use the mathematical morphological operations such as dilation operation, erosion operation, opening operation and closed operation simultaneously, road information is processed, obtains final road information (Fig. 7).