CN106886609B - Block type rural residential area remote sensing quick labeling method - Google Patents
Block type rural residential area remote sensing quick labeling method Download PDFInfo
- Publication number
- CN106886609B CN106886609B CN201710204859.8A CN201710204859A CN106886609B CN 106886609 B CN106886609 B CN 106886609B CN 201710204859 A CN201710204859 A CN 201710204859A CN 106886609 B CN106886609 B CN 106886609B
- Authority
- CN
- China
- Prior art keywords
- image blocks
- image
- residential
- image block
- block
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/10—Image acquisition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/5866—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using information manually generated, e.g. tags, keywords, comments, manually generated location and time information
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/28—Quantising the image, e.g. histogram thresholding for discrimination between background and foreground patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/10—Image acquisition
- G06V10/16—Image acquisition using multiple overlapping images; Image stitching
Abstract
The invention discloses a block type rural residential area remote sensing quick labeling method, which comprises the following steps: acquiring a remote sensing image to be marked, and dividing the remote sensing image into a plurality of image blocks; preliminarily marking the image blocks as residential area image blocks and non-residential area image blocks according to set constraint conditions; acquiring a plurality of preliminarily labeled image blocks of non-residential areas as negative samples and a manually labeled image block as a positive sample, and selecting a set number of image blocks as positive samples by adopting a nearest neighbor classification method; adopting positive samples, negative samples and a plurality of characteristics, and gradually eliminating non-resident image blocks in the resident area image blocks by using a k-nearest neighbor classification method; the image blocks of the rural residential areas with the image blocks adjacent to the image blocks of the rural residential areas in the space are gathered together to form a plurality of image blocks of the gathered residential areas, whether the number of the image blocks of the rural residential areas in the image blocks of the gathered residential areas is smaller than a set value or not is judged, if so, the image blocks of the gathered residential areas are added into a negative sample, and otherwise, corresponding image blocks of the gathered residential areas are output.
Description
Technical Field
The invention relates to a marking technology of residential areas in remote sensing images, in particular to a block type rural residential area remote sensing quick marking method.
Background
With the continuous improvement of the spatial resolution of the remote sensing image, the remote sensing image provides more detailed information such as geometry, texture, shape and the like, the internal composition and the external contour of the residential area can be better displayed on the remote sensing image, and the accurate acquisition of the spatial distribution information of the residential area by using a remote sensing means becomes possible. However, the inside of the residential area is composed of buildings, roads, greenbelts, shadows and the like, the spectral difference of different components is large, the information of the residential area is difficult to extract only by using the spectral information, and the expression form of the residential area on the high-resolution image increases the technical difficulty of the remote sensing automatic extraction algorithm. In order to solve the technical problem and meet the application requirements in the fields of disaster assessment, environmental change, military reconnaissance and the like, researchers at home and abroad research on the residential area information extraction algorithm of the high-resolution remote sensing image.
The existing residential area information extraction algorithm for high-resolution remote sensing images mainly has the following two problems: firstly, an algorithm based on a supervised classification mechanism needs a large number of manually labeled training samples, the labeling process is complicated and time-consuming, and the automation degree of the algorithm cannot reach a practical level; secondly, although the unsupervised mechanism based algorithm does not need training samples, the classification accuracy is relatively low, and further improvement is needed. Therefore, the method overcomes the defects of a supervised classification algorithm and an unsupervised classification algorithm, constructs a rapid extraction algorithm of the information of the residential area of the remote sensing image, which does not need training samples or a small amount of training samples and has high extraction accuracy, and is a problem worthy of in-depth research at present.
Disclosure of Invention
Aiming at the defects in the prior art, the block-type rural residential area remote sensing quick labeling method provided by the invention can accurately detect the residential area in the remote sensing image.
In order to achieve the above object, the present invention adopts the following technical solutions.
The block type rural residential area remote sensing quick labeling method comprises the following steps:
the method comprises the steps of obtaining a remote sensing image to be marked, dividing the remote sensing image into a plurality of image blocks, and enabling adjacent image blocks to be partially overlapped;
preliminarily marking the image blocks as residential area image blocks and non-residential area image blocks according to set constraint conditions;
acquiring a plurality of preliminarily labeled image blocks of non-residential areas as negative samples and a manually labeled image block as a positive sample, and selecting a set number of image blocks as positive samples by adopting a nearest neighbor classification method in an area within a set distance from the positive sample;
adopting a positive sample, a negative sample and a plurality of characteristics, applying a k-nearest neighbor classification method, sequentially utilizing each characteristic training classifier, gradually eliminating non-resident image blocks in the image blocks of the residential areas, and obtaining the image blocks of the rural residential areas after the final characteristic elimination;
the image blocks of the rural residential areas with the image blocks adjacent to the image blocks of the rural residential areas in the space are gathered together to form a plurality of image blocks of the gathered residential areas, whether the number of the image blocks of the rural residential areas in the image blocks of the gathered residential areas is smaller than a set value or not is judged, if so, the image blocks in the area are added into a negative sample, and otherwise, the corresponding image blocks of the gathered residential areas are output.
Preferably, the plurality of characteristics comprise spectral, texture and shape characteristics, when non-residential image blocks in the residential area image blocks are removed step by step, each characteristic is used for removing and preliminarily marking non-residential land in the residential area image blocks according to a set sequence, after the non-residential land is removed by adopting the first two characteristics in the set sequence, the image blocks with the adjacent residential area image blocks are required to be aggregated to form an aggregation block, and the aggregation blocks with the number of the residential areas smaller than a set value in the aggregation block are added into the negative sample; and judging the residential land obtained after the previous feature is removed by adopting the second feature and the third feature in the set sequence.
Further, preferably, the image block is a square image block, and the image block is equally divided into four square sub-image blocks;
and judging whether the number of the image blocks of the four rural residential areas with any sub-image block as the center is larger than or equal to a preset value, if so, taking the sub-image block as the residential area, and outputting the sub-image block as the sub-image block of the residential area.
Further, the specific method for preliminarily labeling the image blocks as residential areas and non-residential areas by adopting the preset constraint conditions comprises the following steps:
calculating the window variance of each pixel in each image block, and judging whether the ratio of the window variances of all the pixels smaller than or equal to a set threshold is larger than or equal to a preset threshold or not; if the image block is larger than or equal to the preset threshold value, the image block is eliminated;
extracting the edges of the remote sensing images of the rest image blocks by adopting a canny operator, judging whether the ratio of the number of continuous edge pixels in the horizontal direction to the number of continuous edge pixels in the vertical direction in each image block is in a set range, and if not, rejecting the image block;
judging whether the whole row and column pixels in the rest image block do not contain the number of edge pixels, if any value of the row number and the column number is larger than half of the size of the image block, rejecting the image block;
calculating a gray level cumulative histogram of each image block in the rest image blocks, judging whether the proportion of the pixel gray level value in a set interval is greater than a set value, and if so, rejecting the image block;
and classifying the removed image blocks into preliminarily labeled image blocks of non-residential areas, and classifying the rest image blocks into preliminarily labeled image blocks of residential areas.
Further, it is preferable that the set interval is [0, d ]1]Or interval [ d2,∞]D is said1And d2The acquisition method comprises the following steps:
calculating a gray level cumulative histogram of the manually marked image block;
d is calculated by the following formula1And d2：
s.t.H(i)＞25％×w2
s.t.H(i)＞75％×w2
Wherein, h (i) is the accumulated pixel number when the gray value of the manually marked image block is i, and w is the side length of the image block.
Further, preferably, the training the classifier by using each feature in turn, and the step-by-step elimination of non-residential image blocks in the residential area image blocks further comprises:
selecting positive samples and negative samples with the same quantity to construct a sample set, and then constructing a plurality of classifiers;
the classifier calculates the image blocks preliminarily marked as the residential areas and the spectrum, texture and shape characteristics of the sample set by using a k-nearest neighbor classification method, and searches k nearest neighbors in front of the image blocks of the single residential area in the sample set;
and if at least half of the classifiers determine that the probability of judging that the current preliminary labeled image block of the residential area is smaller than the probability of judging that the current preliminary labeled image block of the non-residential area in the former k neighbors is smaller, classifying the current preliminary labeled image block of the residential area as the non-residential area.
Preferably, the euclidean distance is adopted in the nearest neighbor classification method to select a set number of image blocks as the positive samples, and the selection process of the positive samples further includes:
taking the gray level histogram as the image block characteristics, measuring the difference between the characteristics by Euclidean distance, and calculating the Euclidean distance between all image blocks and the positive samples and all negative samples within the set distance from the manually marked positive samples;
and if the Euclidean distance from the positive samples in all the Euclidean distances of each image block is the minimum value, marking the Euclidean distances as the positive samples until the number of the positive samples reaches the set number.
Further, preferably, when the positive samples within the set distance from the manually labeled positive sample are smaller than the set number, the gray histogram is used as the image block feature, the euclidean distance between the image block preliminarily labeled as the residential area and the manually labeled positive sample is calculated, and the image block with the smaller euclidean distance is selected to complement the set number of positive samples.
Further, it is preferable that the number Q of positive samples and the number Q of negative samples have a relationship Q ═ nq, and n is singular.
Further, preferably, the remote sensing image is a high-resolution second panchromatic image.
The invention has the beneficial effects that: when the marking of the block-type residential areas is carried out, only one artificially marked sample is needed, a large number of artificially marked data samples are not needed, the complex degree of the marking process is reduced, and the automatic and real-time marking requirements of the block-type rural residential areas are met in the face of sea-level remote sensing image data.
By adopting the method to mark the residents, the residents in the image can be efficiently and accurately detected, and the remote sensing image with difference in ground object background has good effect; compared with the manually marked structure, the marking result of the method has no condition that a large non-residential area is wrongly judged as a residential area, and the precision of manual marking can be achieved on the whole.
Drawings
FIG. 1 is a flow chart of an embodiment of a block-type rural residential area remote sensing quick labeling method.
Fig. 2 is a schematic diagram of various division modes of the remote sensing image.
Fig. 3 is a schematic diagram of overlapped image blocks in two different division modes.
Fig. 4 is a schematic diagram of partial blocks after overlapping image blocks of two different division modes.
Fig. 5 is a schematic diagram of a positive sample search method.
FIG. 6 is a remote sensing image map of a block-type rural residential area marked by the method of the present scheme.
Fig. 7 is a remote sensing image after block-type rural residential area labeling is performed by adopting the existing high-resolution remote sensing image rural residential area extraction algorithm based on edge inhibition point voting.
FIG. 8 is a remote sensing image after block-based rural residents are marked by a manual marking method.
Detailed Description
The following description of the embodiments of the present invention is provided to facilitate the understanding of the present invention by those skilled in the art, but it should be understood that the present invention is not limited to the scope of the embodiments, and it will be apparent to those skilled in the art that various changes may be made without departing from the spirit and scope of the invention as defined and defined in the appended claims, and all matters produced by the invention using the inventive concept are protected.
Referring to fig. 1, fig. 1 shows a flow chart of an embodiment of a block-type rural residential area remote sensing quick labeling method; as shown in fig. 1, the method 100 includes steps 101 to 105.
In step 101, obtaining a remote sensing image to be labeled, dividing the remote sensing image into a plurality of image blocks (the window side length of the image block in the scheme is selected as w pixels), and partially overlapping adjacent image blocks; when the method is implemented, the remote sensing image selected by the scheme is a high-resolution No. two panchromatic image.
In step 102, the image blocks are preliminarily labeled as residential area image blocks and non-residential area image blocks according to the set constraint conditions.
In an embodiment of the present invention, a specific method for preliminarily labeling image blocks as residential areas and non-residential areas by using set constraints (the constraints set in the present scheme include four, namely constraint 1, constraint 2, constraint 3, and constraint 4) is as follows:
constraint 1: calculating the window variance of each pixel in each image block, and judging whether the ratio of the window variances of all the pixels smaller than or equal to a set threshold is larger than or equal to a preset threshold or not; if the number of the image blocks is larger than or equal to the number of the image blocks, the image blocks are eliminated.
Wherein, for the pixel (i, j) at the ith row and the jth column in the image block, the calculation formula of the window variance v (i, j) is as follows:
where W is the side length of the window and g (i, j) is the gray scale value of the pixel (i, j).
The calculation method for setting the threshold value comprises the following steps: selecting a single manually marked image block, and calculating a cumulative histogram H of the window variance of the manually marked image block, wherein the calculation formula of the set threshold value T is as follows:
s.t.H(i)＞10％×w2
where h (i) is the cumulative number of pixels when the gray value is i, and w is the side length of the video block B. The preset threshold is set to be as high as possible, so that the excluded image blocks are further guaranteed not to be residents, the preset threshold in the scheme is set to be 95, and the side length W of the window is set to be 5 in the process of calculating the variance of the window.
Constraint 2: and (3) extracting the edges of the remote sensing image by adopting a canny operator for the rest image blocks (the rest image blocks after constraint condition 1 removal), judging whether the ratio of the number of continuous edge pixels in the horizontal direction to the number of continuous edge pixels in the vertical direction in each image block is in a set range (the set range is [1/R, R ], wherein the value of R is preferably 3.5), and removing the image block if the ratio is not in the set range.
The calculation formula of the number of edge pixels rcn that are continuous in the horizontal direction is:
the formulas of the number ccn of edge pixels continuous in the vertical direction are respectively as follows:
in the formula, e (I, j) is an edge extraction result of the canny operator, if the pixel (I, j) is an edge pixel, the value of e (I, j) is 1, otherwise, the value of e (I, j) is 0, I is an indication function, if the condition is met, the function value is 1, otherwise, the value is 0, and w is the side length of the image block B.
Constraint 3: and judging that the whole row and whole column of pixels in the rest image blocks (the image blocks left after the constraint condition 2 is removed) do not contain the number of edge pixels, and removing the image blocks if any value of the row number and the column number is larger than half of the size of the image blocks.
In practice, the formula for calculating the number of lines rfn for which no edge pixel is included in the entire line of pixels in the remaining image block is:
the formula for calculating the number of rows rfn that no edge pixel is included in the entire row of pixels in the remaining image block is:
in the formula, e (I, j) is the edge extraction result of the canny operator, I is the indicator function, and w is the side length of the image block B.
Constraint 4: and calculating the gray level cumulative histogram of each image block in the rest image blocks (the image blocks left after the constraint condition 3 is removed), judging whether the proportion of the pixel gray level value in the set interval is greater than a set value (the preferred set value in the scheme is selected to be 90), and removing the image block if the proportion is greater than the set value.
In one embodiment of the present invention, the interval is set to [0, d ]1]Or interval [ d2,∞]，d1And d2The acquisition method comprises the following steps:
calculating a gray level cumulative histogram of the manually marked image block;
d is calculated by the following formula1And d2：
s.t.H(i)＞25％×w2
s.t.H(i)＞75％×w2
Wherein, h (i) is the accumulated pixel number when the gray value of the manually marked image block is i, and w is the side length of the image block.
After the residential area and the non-residential area are divided by adopting the four constraint conditions, the removed image blocks are classified into the primarily labeled non-residential area image blocks, and the rest image blocks are classified into the primarily labeled residential area image blocks.
When the constraint conditions are adopted to judge the residential areas and the non-residential areas, the image blocks of the non-residential areas are primarily judged to be directly excluded, and then when the next constraint condition is judged, the rest image blocks are directly judged, namely the image blocks of the non-residential areas, which are excluded from the constraint conditions 1, do not participate in the judgment of the constraint conditions 2, the image blocks of the non-residential areas, which are excluded from the constraint conditions 1 and 2, do not participate in the judgment of the constraint conditions 3, and the image blocks of the non-residential areas, which are excluded from the constraint conditions 1, 2 and 3, do not participate in the judgment of the constraint conditions 4.
In step 103, a plurality of preliminarily labeled image blocks of the non-residential area are obtained as negative samples and a manually labeled image block is obtained as a positive sample, and in an area within a set distance from the positive sample, a set number of image blocks are selected as positive samples by adopting a nearest neighbor classification method;
in the implementation, the present scheme preferably selects a set number of image blocks as positive samples by using the euclidean distance in the nearest neighbor classification method, wherein the selection process of the positive samples further includes:
in the area within the set distance from the positive sample, selecting a set number of image blocks as the positive sample by using the euclidean distance further includes:
calculating Euclidean distances among all the image blocks, manually marked positive samples and all negative samples in a range from a manually marked positive sample set value by adopting a gray level histogram of the image blocks;
and if the Euclidean distance from the positive samples in all the Euclidean distances of each image block is the minimum value, marking the Euclidean distances as the positive samples until the number of the positive samples reaches the set number.
The following description is made after further refinement of the selection of positive samples:
firstly, randomly selecting q image blocks from image blocks divided into non-residential areas as negative samples, and selecting a single image block of the residential area marked by hand as a positive sample; then, performing primary expansion on the positive sample by using a neighborhood image block of the single positive sample; and finally, further selecting positive samples from the image blocks preliminarily divided into the residential areas.
The other positive samples are searched starting from a picture block two picture block sizes away from the current positive sample, as shown in fig. 5. In fig. 5, the black image block is an artificially labeled positive sample, and the gray image block is a starting search image block. And classifying the gray image blocks by using a nearest neighbor classification method on the basis of the current positive sample and the current negative sample. Specifically, a gray histogram h of the image block is calculated as the image block feature, and the Euclidean distance is used as the image block featureTo measure the similarity between histograms, n represents the number of gray levels.
Setting the initial image block BlHas a gray level histogram of hlThe current sample set is S { (h)1,c1),(h2,c2),...,(hQ+1,cQ+1) The first sample is a positive sample, and the class mark c thereof1Is 1, the remaining samples are negative samples, whose class sample is labeled 0. For the initial image block BlCalculating the S-center and h-histogram of the sample setlThe closest sample h 'corresponds to a class c', i.e.:
will start the image block BlThe decision is class c', and if the starting image block is classified as a non-residential area, the image block is moved in the direction of the arrow in fig. 5 until the image block is classified as a residential area. Positive sample selection using this methodAnd the position requirement on the manually marked positive samples is avoided, and certain difference among the positive samples is also ensured.
After the neighborhood positive samples are selected, if the positive samples within a set distance from the manually marked positive samples are smaller than a set number, calculating Euclidean distance between the gray level histogram of the image block in the preliminarily marked residential area and the manually marked positive samples, and selecting the image block with a smaller Euclidean distance to complement the set number of positive samples.
In step 104, a positive sample, a negative sample and a plurality of features are adopted, a k-nearest neighbor classification method is applied, each feature training classifier is sequentially used, non-resident image blocks in the resident area image blocks are gradually eliminated, and rural resident area image blocks with the last feature eliminated are obtained.
During implementation, each feature training classifier is sequentially utilized, and non-resident image blocks in the resident area image blocks are gradually eliminated, so that the method further comprises the following steps:
selecting positive samples and negative samples with the same quantity to construct a sample set, and then constructing a plurality of classifiers;
the classifier calculates the image blocks preliminarily marked as the residential areas and the spectrum, texture and shape characteristics of the sample set by using a k-nearest neighbor classification method, and searches k nearest neighbors in front of the image blocks of the single residential area in the sample set;
and if at least half of the classifiers determine that the probability of judging that the current preliminary labeled image block of the residential area is smaller than the probability of judging that the current preliminary labeled image block of the non-residential area in the former k neighbors is smaller, classifying the current preliminary labeled image block of the residential area as the non-residential area.
The specific algorithm for gradually eliminating the non-resident image blocks in the resident image blocks can be further refined in the following way: after the features (spectrum, texture and SIFT features) of the image blocks are extracted, the spectrum, the texture and the SIFT features are sequentially applied, the image blocks which are preliminarily determined as residential areas are classified under the support of automatically selected positive and negative samples, and the image blocks of non-residential areas are removed.
The method for rejecting the image blocks of the non-residential areas by using the single characteristic comprises the following steps:
(1) and randomly selecting Q image blocks from the non-residential area image blocks as a negative sample set. The number of positive samples Q and the number of negative samples Q satisfy the relationship Q ═ nq, n being singular. Increasing the value of n allows more negative samples to be included in the training samples.
(2) And constructing n classifiers and judging the image blocks of the residential areas. And for the image block B, if the number of the classifiers which are determined as the residential area is more than n/2, determining the image block B as the residential area. When a single classifier is constructed, q negative sample sets are selected from the negative sample sets, and the negative sample sets and q positive samples form a sample set D. Setting image block B to be judgedtIs characterized by htThe current sample set is S { (h)1,c1),(h2,c2),L,(hN,cN) Where N ═ 2 q. For image block BtFind its first k neighbors in the sample set D, where k is set1Individual sample belongs to the inhabitant, k2Individual samples belong to non-populated areas. Number of neighboring samples k if belonging to the residential area1Greater than the number k of neighboring samples belonging to the non-populated area2Then, the image block B is determinedtIs a residential area; otherwise, judging the image block BtIs a non-residential area. In the classifier, the euclidean distance is used to measure the difference between features.
In step 105, clustering rural residential area image blocks in a space with adjacent rural residential area image blocks to form a plurality of clustered residential area image blocks, judging whether the number of the rural residential area image blocks in the clustered residential area image blocks is less than a set value, and if so, adding the image blocks in the area into a negative sample; otherwise, outputting the corresponding image block of the gathered residential area.
When non-residential image blocks in the residential area image blocks are gradually eliminated, adopting each feature elimination according to a set sequence to eliminate the non-residential land preliminarily marked as the residential area in the residential area image blocks, adopting the first two features of the set sequence to eliminate the non-residential areas, gathering the image blocks with the neighboring residential area image blocks to form a gathering block, and adding the gathering blocks of which the number of the residential areas is less than a set value into a negative sample; and judging the residential land obtained after the previous feature is removed by adopting the second feature and the third feature in the set sequence.
The method of residential area aggregation is further described below:
searching adjacent residential areas, generating an aggregated residential area image block R, calculating the number m of regular image blocks contained in the residential area image block R, and adding the area where the image block R is located into a negative sample set if the number m is smaller than a certain threshold value T. As the block type rural residential areas have a certain scale, the residential area is composed of a certain amount of image blocks. The image block R of the large residential area composed of the image blocks B with less number is a non-residential area with high probability, and the smaller the number m is, the higher the probability of belonging to the non-residential area is, and T is 8 in the scheme.
From a single rural residential area image block BtStarting from the image block B of the spatially adjacent rural residential areasnAdded to the image block R of the gathered residential areatThen further from the image block B of the adjacent rural residential areanStarting from the starting point, more image blocks adjacent to rural residential areas are added into the Rt. This process is repeated until RtNo new image block of adjacent residential areas is added;
in this way, starting from the image block of the rural residential area which is not defined to the image block of the current gathered residential area, the image block of the adjacent rural residential area is searched until the image blocks of the adjacent rural residential areas of all the image blocks of the rural residential area are added into the corresponding image block of the gathered residential area;
finally, each of the aggregated residential area image blocks R is calculatedtThe number of the image blocks in the rural residential area is m.
In order to further improve the accuracy that the image block of the rural residential area is the residential area, in the implementation process, the image block is preferably a square image block, and the image block is equally divided into four square sub-image blocks;
and judging whether the number of the image blocks of the four rural residential areas with any sub-image block as the center is larger than or equal to a preset value, if so, taking the sub-image block as the residential area, and outputting the sub-image block as the sub-image block of the residential area.
The following describes the division of image blocks in the remote sensing image:
as shown in fig. 2, the residential areas are labeled in units of square image blocks. Setting the side length of a window as w pixels, and dividing the remote sensing image in an overlapping mode:
and (3) taking the upper left corner of the remote sensing image as an origin, starting from points (0,0), (w/2,0), (0, w/2) and (w/2 ), respectively, moving the window by step length w in the horizontal direction and the vertical direction, and recording an image block B where the window completely falls in the image is located. The manner of dividing the remote sensing image is shown in fig. 2.
The division results in (a) and (d) in fig. 2 are superimposed to obtain the blocking scheme shown in fig. 3. Similarly, the results of the division in (b) and (c) in fig. 2 are superimposed to obtain the blocking scheme shown in fig. 3. Therefore, after the remote sensing image is divided in the above manner, the remote sensing image can be regarded as being divided into w/2 square image blocks b, and there is no overlap between the image blocks b, the image block mentioned in the present embodiment is a w square image block, and the sub-image block is a w/2 square image block.
As shown in FIG. 4, four adjacent sub-image blocks b with side length of w/2i-1,j-1、bi,j-1、bi-1,j、bi,jConstituting a large image block B. Wherein, the sub-image block bijIs four image blocks B1＝{bi-1,j-1、bi,j-1、bi-1,j、bi,j}、B2＝{bi,j-1、bi+1,j-1、bi,j、bi+1,j}、B3＝{bi-1,j、bi,j、bi-1,j+1、bi,j+1}、B4＝{bi,j、bi+1,j、bi,j+1、bi+1,j+1The common part of. If the image block B1、B2、B3、B4If more than 1 block is the image block of the rural residential area, the sub-image block b is usedijAnd judging as a residential area image block.
In one embodiment of the present invention, the spectral feature of the image block is selected from a gray histogram, the texture feature of the image block is selected from a window variance histogram, the histogram intervals have the same size, and the number of the intervals is 64. The image block shape feature selects the selected SIFT (Scale-invariant feature transform) feature. The SIFT algorithm consists of two processes of feature point extraction and feature point description.
When the SIFT feature extraction algorithm is adopted, the process of feature point extraction is omitted, the center point of the subblock with the size of 16 multiplied by 16 is directly taken as the feature point, feature description is carried out on the center point, and a 128-dimensional feature is generated. Due to the adoption of the SIFT feature, the side length of the image block B extracted by the residents is at least 16 pixels. The scheme sets the side length w of the image block B extracted by residents as m times of 16, and then performs non-overlapping division on the image block B into m2Sub-blocks with a side length of 16 × 16 and SIFT features s are extracted. Dimension of a single SIFT feature is 128 if for m2Connecting the 128-dimensional features, the resulting feature dimensions are too large. Let the extracted SIFT feature be si,i＝1,2,…,m2Is as follows foriThe SIFT feature S of the image block is obtained by fusioni：
In the formulaIs a characteristic sjThe ith dimension component of (1).
The following describes the effect of the labeling method in the present scheme with reference to the existing edge-suppression-point-voting-based high-resolution remote sensing image rural residential area extraction algorithm (hereinafter referred to as contrast algorithm) and manual labeling in the prior art:
in the test, the size w of the residential area labeling unit is set to be 32, the number q of positive samples is set to be 50, the number q of negative samples is set to be 450, namely 9 classifiers are trained in the fine classification of the residential areas, and 11 nearest neighbors are selected from each classifier to judge whether the residential areas are the residential areas. The labeling result, the comparison algorithm extraction result and the manual extraction result of the scheme are respectively shown in fig. 6 to 8, and the areas marked by white lines in fig. 6 to 8 are block-type rural residential areas.
As can be seen from fig. 6 to 8, in the face of different background ground object types, the residential area marked by the marking method of the scheme generally conforms to the result of manual marking, and a large non-residential area is not mistakenly judged as a residential area, so that the block-type rural residential area can be effectively extracted from the remote sensing image.
Further adopting 20 remote sensing images to calculate the accuracy and the false alarm rate of the labeling method and the comparison algorithm of the scheme so as to quantitatively evaluate the extraction performance of the algorithm of the scheme, wherein the average accuracy and the false alarm rate of the scheme are 92.34 percent and 2.73 percent respectively; the average accuracy and false alarm rate of the comparison algorithm are 92.52% and 3.52%, respectively.
Therefore, the accuracy rates of the labeling algorithm and the comparison algorithm are in the same level, but the false alarm rate of the labeling algorithm is lower, and fewer non-residential areas in the labeling result are misjudged as residential areas.
Claims (9)
1. The block-type rural residential area remote sensing quick labeling method is characterized by comprising the following steps:
the method comprises the steps of obtaining a remote sensing image to be marked, dividing the remote sensing image into a plurality of image blocks, and enabling adjacent image blocks to be partially overlapped;
preliminarily marking the image blocks as residential area image blocks and non-residential area image blocks according to set constraint conditions;
acquiring a plurality of preliminarily labeled image blocks of non-residential areas as negative samples and a manually labeled image block as a positive sample, and selecting a set number of image blocks as positive samples by adopting a nearest neighbor classification method in an area within a set distance from the positive sample;
adopting a positive sample, a negative sample and a plurality of characteristics, applying a k-nearest neighbor classification method, sequentially utilizing each characteristic training classifier, gradually eliminating non-resident image blocks in the resident area image blocks, and obtaining rural resident area image blocks after the final characteristic elimination, wherein the characteristics comprise spectrum, texture and shape characteristics;
the image blocks of the rural residential areas with the image blocks adjacent to the image blocks of the rural residential areas in the space are gathered together to form a plurality of image blocks of the gathered residential areas, whether the number of the image blocks of the rural residential areas in the image blocks of the gathered residential areas is smaller than a set value or not is judged, if so, the image blocks in the area are added into a negative sample, and otherwise, the corresponding image blocks of the gathered residential areas are output.
2. The remote sensing quick labeling method for the block-type rural residential areas according to claim 1, characterized in that when non-residential image blocks in the residential area image blocks are removed step by step, each feature is adopted to remove the preliminary labeling as non-residential land in the residential area image blocks according to a set sequence, after the non-residential land is removed by adopting the first two features of the set sequence, the image blocks with the neighboring residential area image blocks are aggregated to form an aggregation block, and the aggregation blocks with the number of the residential areas smaller than a set value in the aggregation block are added into a negative sample; and judging the residential land obtained after the previous feature is removed by adopting the second feature and the third feature in the set sequence.
3. The method according to claim 1, wherein the image block is a square image block, and the image block is equally divided into four square sub-image blocks;
and judging whether the number of the image blocks of the four rural residential areas with any sub-image block as the center is larger than or equal to a preset value, if so, taking the sub-image block as the residential area, and outputting the sub-image block as the sub-image block of the residential area.
4. The street-type rural residential area remote sensing quick labeling method according to claim 1, wherein the specific method for preliminarily labeling the image blocks as residential areas and non-residential areas by adopting a preset constraint condition comprises the following steps:
calculating the window variance of each pixel in each image block, judging whether the ratio of the window variances of all the pixels smaller than or equal to a set threshold value is larger than or equal to a preset threshold value or not, and if the ratio is larger than or equal to the preset threshold value, rejecting the image block;
extracting the edges of the remote sensing images of the rest image blocks by adopting a canny operator, judging whether the ratio of the number of continuous edge pixels in the horizontal direction to the number of continuous edge pixels in the vertical direction in each image block is in a set range, and if not, rejecting the image block;
judging whether the whole row and column pixels in the rest image block do not contain the number of edge pixels, if any value of the row number and the column number is larger than half of the size of the image block, rejecting the image block;
calculating a gray level cumulative histogram of each image block in the rest image blocks, judging whether the proportion of the pixel gray level value in a set interval is greater than a set value, and if so, rejecting the image block;
and classifying the removed image blocks into preliminarily labeled image blocks of non-residential areas, and classifying the rest image blocks into preliminarily labeled image blocks of residential areas.
5. The remote sensing quick labeling method for the block-type rural residential areas according to claim 4, wherein the set interval is [0, d ]1]And interval [ d2,∞]D is said1And d2The acquisition method comprises the following steps:
calculating a gray level cumulative histogram of the manually marked image block;
d is calculated by the following formula1And d2：
s.t.H(i)>25％×w2
s.t.H(i)>75％×w2
Wherein, h (i) is the accumulated pixel number when the gray value of the manually marked image block is i, and w is the side length of the image block.
6. The remote sensing quick labeling method for the street type rural residential areas according to claim 1, wherein the step of training the classifier by sequentially utilizing each feature and gradually eliminating the non-residential image blocks in the residential area image blocks comprises the following steps:
selecting positive samples and negative samples with the same quantity to construct a sample set, and then constructing a plurality of classifiers;
the classifier calculates the image blocks preliminarily marked as the residential areas and the spectrum, texture and shape characteristics of the sample set by using a k-nearest neighbor classification method, and searches k nearest neighbors in front of the image blocks of the single residential area in the sample set;
and if at least half of the classifiers determine that the probability of judging that the current preliminary labeled image block of the residential area is smaller than the probability of judging that the current preliminary labeled image block of the non-residential area in the former k neighbors is smaller, classifying the current preliminary labeled image block of the residential area as the non-residential area.
7. The remote sensing quick labeling method for the street type rural residential areas according to claim 1, wherein a set number of image blocks are selected as positive samples by using Euclidean distance in the nearest neighbor classification method, and the selection process of the positive samples further comprises:
taking the gray level histogram as the image block characteristics, measuring the difference between the characteristics by Euclidean distance, and calculating the Euclidean distance between all image blocks and the positive samples and all negative samples within the set distance from the manually marked positive samples;
and if the Euclidean distance from the positive samples in all the Euclidean distances of each image block is the minimum value, marking the Euclidean distances as the positive samples until the number of the positive samples reaches the set number.
8. The remote sensing quick labeling method for the neighborhood type rural residential areas according to any one of claims 1 to 7, wherein the relationship between the number Q of the positive samples and the number Q of the negative samples is Q ═ nq, and n is singular.
9. The remote sensing method for rapid labeling of neighborhood-based rural residential areas according to any one of claims 1-7, wherein the remote sensing image is a high-resolution two-color full color image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710204859.8A CN106886609B (en) | 2017-03-31 | 2017-03-31 | Block type rural residential area remote sensing quick labeling method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710204859.8A CN106886609B (en) | 2017-03-31 | 2017-03-31 | Block type rural residential area remote sensing quick labeling method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106886609A CN106886609A (en) | 2017-06-23 |
| CN106886609B true CN106886609B (en) | 2019-12-31 |
Family
ID=59182488
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710204859.8A Expired - Fee Related CN106886609B (en) | 2017-03-31 | 2017-03-31 | Block type rural residential area remote sensing quick labeling method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106886609B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108108721A (en) * | 2018-01-09 | 2018-06-01 | 北京市遥感信息研究所 | A kind of method that road extraction is carried out using EO-1 hyperion |
| CN115035406B (en) * | 2022-06-08 | 2023-08-04 | 中国科学院空间应用工程与技术中心 | Remote sensing scene data set labeling method, remote sensing scene data set labeling system, storage medium and electronic equipment |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102073867A (en) * | 2010-12-27 | 2011-05-25 | 北京师范大学 | Sorting method and device for remote sensing images |
| CN102496034A (en) * | 2011-11-29 | 2012-06-13 | 南京师范大学 | High-spatial resolution remote-sensing image bag-of-word classification method based on linear words |
| CN103049751A (en) * | 2013-01-24 | 2013-04-17 | 苏州大学 | Improved weighting region matching high-altitude video pedestrian recognizing method |
| CN104866852A (en) * | 2015-04-01 | 2015-08-26 | 环境保护部卫星环境应用中心 | Method and apparatus for extracting land cover information in remote sensing image |
| CN105260738A (en) * | 2015-09-15 | 2016-01-20 | 武汉大学 | Method and system for detecting change of high-resolution remote sensing image based on active learning |
| CN105809177A (en) * | 2014-12-31 | 2016-07-27 | 中国科学院深圳先进技术研究院 | Method used for actuating remote sensing image classification |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| TWI478074B (en) * | 2010-12-01 | 2015-03-21 | Inst Information Industry | Code recoginizing method, device and computer readable storage medium for storing code recoginizing method |
-
2017
- 2017-03-31 CN CN201710204859.8A patent/CN106886609B/en not_active Expired - Fee Related
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102073867A (en) * | 2010-12-27 | 2011-05-25 | 北京师范大学 | Sorting method and device for remote sensing images |
| CN102496034A (en) * | 2011-11-29 | 2012-06-13 | 南京师范大学 | High-spatial resolution remote-sensing image bag-of-word classification method based on linear words |
| CN103049751A (en) * | 2013-01-24 | 2013-04-17 | 苏州大学 | Improved weighting region matching high-altitude video pedestrian recognizing method |
| CN105809177A (en) * | 2014-12-31 | 2016-07-27 | 中国科学院深圳先进技术研究院 | Method used for actuating remote sensing image classification |
| CN104866852A (en) * | 2015-04-01 | 2015-08-26 | 环境保护部卫星环境应用中心 | Method and apparatus for extracting land cover information in remote sensing image |
| CN105260738A (en) * | 2015-09-15 | 2016-01-20 | 武汉大学 | Method and system for detecting change of high-resolution remote sensing image based on active learning |
Non-Patent Citations (3)
| Title |
|---|
| Knowledge-based region labeling for remote sensing image interpretation;G. Forestier等;《Computers, Environment and Urban Systems》;20120930;第36卷(第5期);470-480 * |
| 基于主题模型的高空间分辨率遥感影像分类研究;徐盛;《中国博士学位论文全文数据库 信息科技辑》;20120715(第7期);I140-112 * |
| 基于样本自动扩充的街区式农村居民地遥感提取方法;陆尘等;《地球信息科学学报》;20180925;第20卷(第9期);1306-1315 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106886609A (en) | 2017-06-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106845408B (en) | Street garbage identification method under complex environment | |
| CN109903331B (en) | Convolutional neural network target detection method based on RGB-D camera | |
| CN110378308B (en) | Improved port SAR image near-shore ship detection method based on fast R-CNN | |
| CN102426649B (en) | Simple steel seal digital automatic identification method with high accuracy rate | |
| CN110084241B (en) | Automatic ammeter reading method based on image recognition | |
| CN104978567B (en) | Vehicle checking method based on scene classification | |
| CN107492094A (en) | A kind of unmanned plane visible detection method of high voltage line insulator | |
| CN108491498B (en) | Bayonet image target searching method based on multi-feature detection | |
| CN110033431B (en) | Non-contact detection device and detection method for detecting corrosion area on surface of steel bridge | |
| CN112149543B (en) | Building dust recognition system and method based on computer vision | |
| CN111027446B (en) | Coastline automatic extraction method of high-resolution image | |
| CN103530638A (en) | Method for matching pedestrians under multiple cameras | |
| CN108509950B (en) | Railway contact net support number plate detection and identification method based on probability feature weighted fusion | |
| CN107944354B (en) | Vehicle detection method based on deep learning | |
| CN104657980A (en) | Improved multi-channel image partitioning algorithm based on Meanshift | |
| CN110866430A (en) | License plate recognition method and device | |
| CN109635733B (en) | Parking lot and vehicle target detection method based on visual saliency and queue correction | |
| CN111310756A (en) | Damaged corn particle detection and classification method based on deep learning | |
| CN108009556A (en) | A kind of floater in river detection method based on fixed point graphical analysis | |
| CN111274964B (en) | Detection method for analyzing water surface pollutants based on visual saliency of unmanned aerial vehicle | |
| CN110852207A (en) | Blue roof building extraction method based on object-oriented image classification technology | |
| Somawirata et al. | Road detection based on the color space and cluster connecting | |
| CN115035182B (en) | Landslide hazard early warning method and system | |
| CN104484679B (en) | Non- standard rifle shooting warhead mark image automatic identifying method | |
| CN106886609B (en) | Block type rural residential area remote sensing quick labeling method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20191231 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |