CN104951785A - Method for evaluating feature description processes - Google Patents

Abstract

The invention relates to the field of object detection and identification in the field of image identification, in particular to a method for evaluating feature description processes. By the aid of the method for evaluating the feature description processes, problems in the prior art can be solved, the multiple feature description processes are quantitatively compared to one another, and accordingly the optimal feature description processes can be selected. The method has the advantages that to-be-identified objects are sampled to obtain positive samples and negative samples, the positive samples and the negative samples are converted into positive feature description vectors and negative feature description vectors by the aid of the to-be-evaluated feature description processes, then the repeatability of positive-positive feature description vectors and the discrimination of positive-negative feature vectors are quantitatively computed by the aid of clustering analysis processes, and accordingly the to-be-identified object detection and identification performance of the to-be-evaluated feature description processes can be quantitatively assessed.

Description

A method for measuring characterization methods

technical field

The invention relates to the field of object detection and recognition in the field of image recognition, in particular to a method for measuring a feature description method.

Background technique

In the field of object image detection and recognition, the current research hotspot is generic object (Generic ObjectCategory) detection and recognition. Generic objects refer to a class of objects, and there are commonality and individuality between individuals in the same object, such as airplanes, apples, and people. Generic objects are for specific objects. A specific object refers to a particular individual, or a group of people with exactly the same appearance, such as the Eiffel Tower, my bicycle, and a brand new iphone6. compared to specific object detection.

The difficulty of detecting and identifying generic objects is significantly greater than that of specific objects. First of all, the number of individuals contained in a class of objects may be infinite, and it is impossible to use all samples to train the detector, only some representative individuals can be selected. Second, similar objects have common similarities, but these similarities may be abstract, such as smooth, bulging, metallic texture, etc. Finally, different individuals of the same kind of objects often show differences in appearance. How to define the range of differences that a class of objects can contain, and how to further divide them into subcategories based on these differences is also a difficult point.

The design of generic object detectors basically starts with the design or selection of feature description methods. Generally, in the detector training phase, the window sampled in the training image is first converted into a feature vector using a feature description method, and then a variety of machine learning methods are used to train the high-level recognition model and parameters. In the detection and recognition stage, the detection window is first converted into a feature vector using the feature description method, and then the feature vector with a high similarity to the detected object is screened and combined to realize object detection and recognition. It can be seen that the feature description method is the basis of object detector design, and the selection of an appropriate feature description method has a great influence on the overall performance of the detector.

Designing detectors requires a choice of an appropriate characterization method, but previous characterization evaluation methods all have different limitations. The feature description evaluation method commonly used in specific object detection uses two images containing the same specific object to calculate the correct rate of feature point/region matching results through feature vectors. However, the generic objects are not exactly the same, and the exact matching feature points/feature vectors cannot be established. In the detection of generic objects, a measurement system that combines multiple feature description methods with multiple upper-level recognition models is often used. However, such a measurement method will be affected by the upper-level recognition model, and its calculation results are biased.

Contents of the invention

The technical problem to be solved by the present invention is to provide a method for measuring the feature description method in view of the above-mentioned problems. The difference (distinguishability) of features in objects, quantitative comparison of multiple feature description methods, so as to select the best feature description method.

The technical scheme that the present invention adopts is as follows:

One measure of characterization methods includes:

Step 1: Sampling is uniformly distributed through the sampling window for all marked areas of the positive sample image; uniformly distributed sampling is performed for the image of the negative sample through the sampling window;

Step 2: Using the feature description method to be measured, convert all sampling windows of all marked areas of the positive sample image and all sampling windows of the negative sample image into positive feature vectors F _P ={α ₁ ,α ₂ ...α _A } and negative eigenvectors F _N ={β ₁ ,β ₂ ...β _B }, the numbers of which are denoted as A and B respectively;

Step 3: Use a cluster analysis method to divide the positive feature vectors into K clusters {w ₁ ,w ₂ ...w _K } according to the spatial distribution distance of the positive feature vectors, and w _k in each cluster Call it a common descriptor, 1≤k≤K, each common descriptor w _k is a set of n positive feature vectors α _i , w _k is a subset of F _P , n≤A, 10≤K≤1000; If the number of feature vectors n≤X contained in a common descriptor, delete the common descriptor; otherwise, keep the common descriptor; where X is 1% to 50% of the number of positive samples;

Step 4: For the general description vocabulary w _k , use a probability distribution model to calculate the parameter P(f|w _k ) of the probability distribution density formula, and then use all positive feature vectors α _i ∈ F _P , i=1...A Computes the positive discriminative value for this common descriptor Use all the negative feature vectors β _j ∈ F _N , j=1...B to calculate the negative discriminative value of the general descriptor $V_N\left(w_k\right)=-\frac{1}{B}\underset{{\mathrm{\β}}_j\∈f_N}{\mathrm{\Σ}}P\left({\mathrm{\β}}_j\right|w_k);$

Step 5: Calculate the distinguishable value of the general description vocabulary V(w _k )=V _P (w _k )+V _N (w _k );

Step 6: Select M discriminative values with the highest distinguishable values in all common descriptors w _k as key descriptors T={t ₁ ,t ₂ ...t _M }, the discriminative degree of these M key descriptors is sum is the generic object detection performance index of this feature description method, A larger value indicates that the measured feature description method is more suitable for the feature description of the tested object, where m=1...M.

Further, the sampling window in step 1 is a rectangular area in the sample image.

A method for measuring the feature description method also includes performing multi-scale scaling on all the identified regions of the positive sample image and the negative sample image, and in the minimum zoom level of the multi-scale scaling, the sampling window is required by the feature description method to be evaluated Minimum area; at the maximum zoom level of multiple zoom scales, the sampling window of positive samples is the smallest area that includes the entire image identification area, and the sampling window of negative samples is the largest area that does not exceed the image boundary; the sampling window of multiple zoom scales starts from the smallest area The zoom level to the maximum zoom level increases in equal proportions, and the ratio is selected between 1.1 and 2 times; under the same zoom scale, the sampling windows are evenly distributed, and there is a 5% to 95% overlapping area between adjacent sampling windows.

Further, a cluster analysis method in the step 3 includes division method, hierarchical method, graph theory method, grid method, K-Means or Mean Shift.

Further, a probability distribution model in step 4 includes Gaussian distribution, Bernoulli distribution, binomial distribution or Poisson distribution.

In summary, owing to adopting above-mentioned technical scheme, the beneficial effect of the present invention is:

1. The present invention proposes an open platform, and all eigenvector-based feature description methods can be compared fairly by this method; the present invention is applicable to the object detection/recognizer design stage to select feature description methods.

2. The present invention proposes an independent measurement method that does not rely on the upper-level recognition model, and a method for evaluating the feature description method in terms of repeatability and distinguishability;

3. The present invention is targeted to the object to be detected, fully reflects the performance changes of the same description method under different application backgrounds, and can make a targeted selection according to the application background;

4. The use of multi-scale sample sampling windows fully ensures that the detection performance of the feature vector is not affected by the scale.

Detailed ways

All features disclosed in this specification, or steps in all methods or processes disclosed, may be combined in any manner, except for mutually exclusive features and/or steps.

Any feature disclosed in this specification (including any appended claims, abstract), unless otherwise stated, may be replaced by alternative features which are equivalent or serve a similar purpose. That is, unless expressly stated otherwise, each feature is one example only of a series of equivalent or similar features.

Relevant description of the present invention:

The use of the method of the present invention requires a sufficiently large and manually marked training sample library. In the sample library, no less than 30 images containing detected or recognized objects should be used as positive samples, and no less than 5 times the number of positive samples of positive samples and images that do not contain detected or recognized objects should be used as negative samples. The ideal number of positive samples is more than 1000, and the number of negative samples is more than 10 times that of positive samples. The region where the detected or recognized object is located in the positive sample is marked. The positive samples contain one or more detected or recognized objects, and the negative samples do not contain detected or recognized objects, so as to calculate the overall recognition performance of the measured method for one or more detected or recognized objects.

Principle: The present invention samples the positive and negative samples of the object to be recognized, and uses the feature description method to be measured (different feature description methods make the images of the positive samples all identify areas and the images of the negative samples correspond to generate different feature vectors) to convert the positive and negative samples. Negative sampling is positive and negative feature description vectors, and then quantitatively calculates the repeatability of positive-positive feature description vectors and the distinguishability of positive-negative feature vectors through cluster analysis methods, so as to quantitatively evaluate the treatment of feature description methods to be measured The ability to recognize objects for detection and recognition.

In computer vision and machine learning, the feature vector of an image refers to a digitized vector. The vector is calculated from the entire image or a local area, which can characterize the characteristics of the corresponding area, and the amount of data is much smaller than that of the corresponding area. The method of converting an image region into a feature vector is called a feature description method, also called a feature transformation method. For example, the SIFT feature description method can convert a 64×64 or 128×128 pixel image area into a 64×1 feature vector. The sobel feature description method can convert a 3X3 or 9X9 pixel image area into a 2X1 feature vector.

Embodiment one:

Step 1: Sampling is uniformly distributed through the sampling window for all marked areas of the positive sample image; uniformly distributed sampling is performed for the image of the negative sample through the sampling window;

Step 2: Using the feature description method to be measured, convert all sampling windows of all marked areas of the positive sample image and all sampling windows of the negative sample image into positive feature vectors F _P ={α ₁ ,α ₂ ...α _A } and negative eigenvectors F _N ={β ₁ ,β ₂ ...β _B }, the numbers of which are denoted as A and B respectively;

Step 3: Use a cluster analysis method to divide the positive feature vectors into K clusters {w ₁ ,w ₂ ...w _K } according to the spatial distribution distance of the positive feature vectors, and w _k in each cluster Call it a common descriptor, 1≤k≤K, each common descriptor w _k is a set of n positive feature vectors α _i , w _k is a subset of F _P , n≤A, 10≤K≤1000; If the number of feature vectors n≤X contained in a common descriptor, delete the common descriptor; otherwise, keep the common descriptor; where X is 1% to 50% of the number of positive samples;

Step 4: For the general description vocabulary w _k , use a probability distribution model to calculate the parameter P(f|w _k ) of the probability distribution density formula, and then use all positive feature vectors α _i ∈ F _P , i=1...A Computes the positive discriminative value for this common descriptor Use all the negative feature vectors β _j ∈ F _N , j=1...B to calculate the negative discriminative value of the general descriptor $V_N\left(w_k\right)=-\frac{1}{B}\underset{{\mathrm{\β}}_j\∈f_N}{\mathrm{\Σ}}P\left({\mathrm{\β}}_j\right|w_k);$

Step 5: Calculate the distinguishable value of the general description vocabulary V(w _k )=V _P (w _k )+V _N (w _k );

Step 6: Select M discriminative values with the highest distinguishable values in all common descriptors w _k as key descriptors T={t ₁ ,t ₂ ...t _M }, the discriminative degree of these M key descriptors is sum is the generic object detection performance index of this feature description method, A larger value indicates that the measured feature description method is more suitable for the feature description of the tested object, where m=1...M.

Among them, the formula of probability distribution density using multi-dimensional Gaussian distribution for general description vocabulary w _k is:

$\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\sqrt{{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; D.</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}<span\; class="notranslate">\; |</span>}}\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; )</span>}^{<span\; class="notranslate">\; \text{'}</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>$

Among them, the distribution parameter μ is the sample mean value, Σ is the sample covariance matrix, which is obtained through statistical calculation of all feature vectors w _k ={f ₁ ,f ₂ ... f _n } in the common descriptor, and D is the feature vector The dimension of , the symbol ' is a transposition operation; then the positive distinguishable value of the general description vocabulary w _k is

${\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; A</span>}}\underset{{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\sqrt{{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; D.</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}<span\; class="notranslate">\; |</span>}}\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; )</span>}^{<span\; class="notranslate">\; \text{'}</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>$

Negative distinguishable values are:

${\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; B</span>}}\underset{{\mathrm{<span\; class="notranslate">\; \&beta;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; N</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\sqrt{{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; D.</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}<span\; class="notranslate">\; |</span>}}\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&beta;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; )</span>}^{<span\; class="notranslate">\; \text{'}</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&beta;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; .</span>$

Further, the positive samples are sampled, and the sizes of the rectangular window areas of each zoom level are 11×11 pixels, 23×23 pixels, 47×47 pixels, 95×95 pixels, and 191×191 pixels; the negative samples are sampled, The size of the rectangular window area at each zoom level is 11×11 pixels, 23×23 pixels, 47×47 pixels, 95×95 pixels, 191×191 pixels, and 383×383 pixels in sequence. Among them, 11 × 11 pixels are the minimum window area required by the feature description method, 191 × 191 pixels are the minimum window area that can contain the image identification area, 383 × 383 pixels are the largest window area that does not exceed the image boundary, and the zoom increment ratio is 2.

The present invention is not limited to the foregoing specific embodiments. The present invention extends to any new feature or any new combination disclosed in this specification, and any new method or process step or any new combination disclosed.

Claims (5)

1. A method for measuring a characterization method, characterized in that it comprises: Step 1: Sampling is uniformly distributed through the sampling window for all marked areas of the positive sample image; uniformly distributed sampling is performed for the image of the negative sample through the sampling window; Step 2: Using the feature description method to be measured, convert all sampling windows of all marked areas of the positive sample image and all sampling windows of the negative sample image into positive feature vectors F _P ={α ₁ ,α ₂ ...α _A } and negative eigenvectors F _N ={β ₁ ,β ₂ ...β _B }, the numbers of which are denoted as A and B respectively; Step 3: Use a cluster analysis method to divide the positive feature vectors into K clusters {w ₁ ,w ₂ ...w _K } according to the spatial distribution distance of the positive feature vectors, and w _k in each cluster Call it a common descriptor, 1≤k≤K, each common descriptor w _k is a set of n positive feature vectors α _i , w _k is a subset of F _P , n≤A, 10≤K≤1000; If the number of feature vectors n≤X contained in a common descriptor, delete the common descriptor; otherwise, keep the common descriptor; where X is 1% to 50% of the number of positive samples; Step 4: For the general description vocabulary w _k , use a probability distribution model to calculate the parameter P(f|w _k ) of the probability distribution density formula, and then use all positive feature vectors α _i ∈ F _P , i=1...A Computes the positive discriminative value for this common descriptor Use all the negative feature vectors β _j ∈ F _N , j=1...B to calculate the negative discriminative value of the general descriptor Step 5: Calculate the distinguishable value of the general description vocabulary V(w _k )=V _P (w _k )+V _N (w _k ); Step 6: Select M discriminative values with the highest distinguishable values in all common descriptors w _k as key descriptors T={t ₁ ,t ₂ ...t _M }, the discriminative degree of these M key descriptors is sum is the generic object detection performance index of this feature description method, A larger value indicates that the measured feature description method is more suitable for the feature description of the tested object, where m=1...M. 2. A method for measuring the feature description method according to claim 1, characterized in that the sampling window in the step 1 is a rectangular area in the sample image. 3. A method for measuring the feature description method according to claim 1, further comprising multi-scale scaling of all the identified areas of the positive sample image and the negative sample image, and the minimum scaling of the multi-scale scaling is In the level, the sampling window is the minimum area required by the feature description method to be evaluated; in the maximum zoom level of multiple zoom scales, the sampling window of the positive sample is the smallest area that includes the entire image identification area, and the sampling window of the negative sample is not beyond the image The largest area of the boundary; multi-scale sampling windows increase in equal proportions from the minimum zoom level to the maximum zoom level, and the ratio is between 1.1 and 2 times; under the same zoom scale, the sampling windows are evenly distributed, and adjacent sampling windows Between 5% and 95% overlap. 4. a kind of method that character description method is weighed according to claim 1, it is characterized in that a kind of clustering analysis method comprises division method, hierarchy method, graph theory method, grid method, K in the described step 3 -Means or Mean Shift. 5. A method for measuring the feature description method according to claim 1, characterized in that a probability distribution model in said step 4 comprises Gaussian distribution, Bernoulli distribution, binomial distribution or Poisson distribution.