CN106548180A - A kind of method for obtaining the Feature Descriptor for obscuring constant image - Google Patents

Abstract

The invention discloses a method for obtaining a feature descriptor of a fuzzy invariant image. Firstly, the image is divided into a plurality of local blocks, and the subdivision of the local blocks is densely sampled to obtain LPQ+; LPQ+ merge to obtain the feature description of the local block, and then perform FV (Fisher Vector) encoding on all the feature descriptors of the local block and perform corresponding regularization processing to obtain a higher-dimensional fuzzy image feature description. The method for obtaining feature descriptors of fuzzy invariant images provided by the present invention is a more efficient and accurate fuzzy image recognition feature descriptor LPQ+ proposed based on LPQ for fuzzy image recognition, which has high discriminative and relatively small feature dimensions. Compared with the combination of FV coding and traditional descriptors, the combination of FV coding and LPQ+ has better results in recognition accuracy and recognition efficiency.

Description

A Method of Obtaining Feature Descriptors of Fuzzy Invariant Images

technical field

The invention belongs to the technical field of digital image recognition, and more specifically relates to a method for obtaining a feature descriptor of a fuzzy invariant image.

Background technique

In the process of image acquisition, the interference of the environment, the swaying of the camera itself or the loss of focus can easily cause image blurring, which brings practical challenges to image recognition tasks such as face recognition, texture classification and object detection.

For the problem of fuzzy image recognition, the existing technology includes the following two solutions; one is to first deblur the image, and then use traditional descriptors (such as LBP, SIFT and HOG, etc.) to carry out target recognition. The method is suitable for the situation where the fuzzy function is known, and the non-blind deconvolution method is used to deblur the image; the other is to directly identify the fuzzy invariant features of the image extraction, and LPQ (Local Phase Quantization) is The representative one uses the phase information of the image, decorrelates and quantizes it and projects it into an eight-dimensional space, so as to construct a histogram on the entire image and extract a feature descriptor that is invariant to the central symmetric blur.

The above method has the following defects. For the first method, the blur function is usually unknown, so the time overhead of deblurring processing is very large. The difficulty of the optimization problem of PSF (Point Spread Function) estimation under this method is that the image The choice of fuzzy type; LPQ in the second method cannot balance efficiency and accuracy well when faced with complex visual recognition tasks; the existing technology combines LPQ with FV (Fisher Vector) to encode LPQ features to form a Higher-dimensional descriptors are used to represent images to improve the comprehensive performance of LPQ, but the increase in dimension not only increases the experimental overhead, but also increases the demand for training samples to prevent model overfitting.

Contents of the invention

Aiming at the above defects or improvement needs of the prior art, the present invention provides a method for obtaining feature descriptors of fuzzy invariant images. Obtain a higher-dimensional fuzzy image feature description to improve the recognition accuracy and efficiency of fuzzy images.

In order to achieve the above object, according to one aspect of the present invention, a method for obtaining a feature descriptor of a blur-invariant image is provided, comprising the following steps:

(1) Blur the samples in the source image data set, and grayscale the blurred image;

(2) Using a sliding window to densely sample the image obtained by grayscale processing to obtain a plurality of local image blocks, and divide each local image block into a plurality of unit blocks;

(3) extracting LPQ+ for each unit block, and then merging the extracted multiple groups of LPQ+ to obtain the feature descriptor of the local block;

Among them, LPQ+ is a feature descriptor based on LPQ, and LPQ+ is obtained by directly quantizing the phase rather than the real part and imaginary part (essentially STFT coefficient) described in LPQ;

(4) Perform Fisher Vector encoding, power regularization and Regularization processing to obtain feature descriptors of blurred images.

Preferably, the above-mentioned method for obtaining the feature descriptor of the fuzzy invariant image also includes the step of identification after the step (4):

(5) According to the training data set and the corresponding feature descriptor, a one-to-one multi-class classification strategy is used to train a linear support vector machine classifier; the output result of the classifier is the recognition result.

Preferably, the above-mentioned method for obtaining the feature descriptor of the fuzzy invariant image, its step (2) includes the following sub-steps:

(2-1) Using a sliding window with a fixed size and a step size to densely sample the image obtained by grayscale processing to obtain K groups of local image blocks;

(2-2) Divide the local image block into multiple small units for storing the spatial structure information of the local block.

Preferably, the above-mentioned method for obtaining the feature descriptor of the blur-invariant image, its step (3) includes the following sub-steps:

(3-1) Using STFT (Short Term Fourier Transform), the phase information is extracted for each pixel in the unit block with the M×M size neighborhood around the central pixel as the calculation range;

Among them, (x, y), (u, v) refer to the coordinates in the fuzzy image and the corresponding Fourier transform domain respectively; g(x, y) refers to the fuzzy image, and G(u, v) is its corresponding Fourier transform form ; N _x and N _y refer to the neighborhood range of the pixel point (x, y).

Take four low-frequency points, u ₁ =(a,0), u ₂ =(0,a), u ₃ =(a,a), u ₄ =(a,-a), and perform STFT on them to obtain Four phase values ∠G(u ₁ ), ∠G(u ₂ ), ∠G(u ₃ ), ∠G(u ₄ ); where a is a constant;

(3-2) Press The phase value is divided to obtain 1 different angle areas, and the above four phase values are classified into the corresponding angle area and quantified according to the following voting values,

Among them, α is an adjustable parameter, Refers to the voting value obtained by ∠G(u _j ) under the corresponding angle area i; where, j=1,2,3,4;

(3-3) Express the distribution of quantized values of the four phase values as four histograms and merge the histograms to obtain LPQ+;

(3-4) Merge the LPQ+ of different unit blocks of the same block to obtain the feature descriptor of the local block.

Preferably, the above-mentioned method for obtaining the feature descriptor of the fuzzy invariant image, its step (4) includes the following sub-steps:

(4-1) Denote the extracted K local block feature descriptors as X={x _k ,k=1,2,...,K};

(4-2) Use N-ary mixed Gaussian model u _λ (x) to simulate the generation process of X, which is denoted as

Among them, the parameter λ={ω _l ,μ _l ,σ _l ,l=1,…,N}, ω _l , μ _l and σ _l refer to the mixing weight, mean vector and standard deviation of the Gaussian function u _l respectively; the parameter λ is estimated on local feature descriptors generated from a large number of training sets according to the EM criterion (expectation maximum criterion);

(4-3) Calculate partial derivatives of Gaussian model u _λ (x) for μ _l and σ _l respectively, and obtain two gradient vectors:

Among them, γ _k (l) represents the probability that the feature x _k is generated by the l-th Gaussian function;

(4-4) Merge all N Gaussian function corresponding gradient vectors to obtain Fisher Vector code

(4-5) right Each dimension m of is subjected to the following power normalization (power normalization, processing:

(4-6) After the above power regularization conduct Regularization processing, get the descriptor of fuzzy image features:

Generally speaking, compared with the prior art, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

(1) The method for obtaining the feature descriptor of the fuzzy invariant image provided by the present invention is based on LPQ, but directly quantizes the real part and the imaginary part (essentially STFT coefficients) described in the phase instead of LPQ to obtain LPQ+; LPQ essentially only uses the symbol information of the STFT coefficients and ignores the specific values, while LPQ+ can make better use of the discriminative properties of LPQ in order to compensate for the quantization of specific phase values; at the same time, because the LPQ+ feature dimension is lower than that of LPQ, Therefore, it has the effect of improving processing efficiency;

(2) The method for obtaining the feature descriptors of fuzzy invariant images provided by the present invention uses Fisher Vector encoding for local feature descriptions, and projects low-dimensional feature descriptions to high-dimensional feature spaces, combined with the extremely large linear support vector machine classifier Improve the description performance of the descriptor and the recognition accuracy of blurred images;

Therefore, the present invention provides a method for obtaining feature descriptors of blur-invariant images with higher recognition accuracy and higher speed.

Description of drawings

FIG. 1 is a schematic flowchart of a method for obtaining a feature descriptor of a blur-invariant image provided by an embodiment of the present invention;

Fig. 2 is a schematic flow chart of extracting feature descriptor LPQ+ in an embodiment of the present invention;

FIG. 3 is a schematic flow diagram of obtaining a local feature descriptor by densely extracting LPQ+ in an embodiment of the present invention;

Fig. 4 is a schematic flow diagram of encoding local block descriptors to form image feature descriptors by using Fisher Vector in an embodiment of the present invention.

detailed description

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not constitute a conflict with each other.

The present invention provides a fuzzy invariant image feature description, the process of which is shown in Figure 1, including the acquisition of fuzzy images, intensive sampling of images, extraction of LPQ+, encoding with Fisher Vector, and training of support vector machine linear classifiers , to obtain the recognition result; the method for obtaining the feature descriptor of the fuzzy invariant image provided by the present invention is specifically described below in conjunction with the embodiments, and the specific steps are as follows:

Step 1, obtaining blurred image samples, including the following sub-steps:

(1-1) Acquire the source image data set; in the present embodiment, five groups of different types of data sets are adopted: the Yale face data set of 15 categories, the KTH texture data set of 10 categories, and the land use scene data set of 21 categories , 6 categories of HUST cloud datasets and 17 categories of Oxford Flower datasets;

(1-2) Acquire blurred images; in the embodiment, three image blurring methods are adopted: Gaussian blur, motion blur and circular blur; specifically, for Gaussian blur, the size of the neighborhood window is set to 3×3, and the standard deviation Set to 1, 2, 3 respectively; for motion blur, the linear motion parameters of the camera are set to 8, 9 respectively; for circular blur, the radius is set to 2, 3 respectively;

(1-3) Grayscale all color images according to the following formula:

K=0.2989×R+0.5870×G+0.1140×B;

Among them, K is the grayscale image, and R, G, and B are the three channels of the color image.

Step 2, densely sample the image, including the following sub-steps:

(2-1) Using a sliding window with a size of 16×16 and a horizontal step size and a vertical step size of 8 to densely sample the image obtained by grayscale processing to obtain a local image block;

(2-2) Divide the acquired local image block into 4 unit blocks, which are used to store the spatial structure information of the local block.

Step 3, extracting LPQ+ from blocks of different units, and then merging them into feature descriptors of local blocks; wherein, the flow charts of extraction of feature descriptors of local blocks and dense sampling are shown in Figure 2 and Figure 3 respectively; Specific steps are as follows:

(3-1) Use STFT (Short Term Fourier Transform) to extract phase information for each pixel in the unit; the calculation range of STFT is a 13×13 neighborhood around the central pixel:

Among them, (x, y), (u, v) refer to the coordinates in the fuzzy image and the corresponding Fourier transform domain respectively; g(x, y) refers to the fuzzy image, and G(u, v) is its corresponding Fourier transform form ; N _x and N _y refer to the neighborhood range of the pixel point (x, y).

Taking four low-frequency points, u ₁ =(a,0), u ₂ =(0,a), u ₃ =(a,a), u ₄ =(a,-a), where the parameter And perform STFT processing on it to get four phase values ∠G(u ₁ ), ∠G(u ₂ ), ∠G(u ₃ ), ∠G(u ₄ );

(3-2) According to Divide the phase value into 8 different angle areas; classify the above four phase values into the corresponding angle area and quantify according to the following voting values,

Refers to the voting value obtained by ∠G(u _j ) under the corresponding angle area i; where, j=1,2,3,4;

(3-3) Expressing the distribution of the quantized values of the four phases as four histograms and merging the histograms to obtain LPQ+;

Step 4, perform Fisher Vector encoding processing, power regularization and Regularization processing obtains the feature descriptor of the fuzzy image; in the embodiment, the realization of Fisher Vector is based on VLFeat, and its flow process is shown in Figure 4, including the following sub-steps:

(4-1) Denote the extracted K local block feature descriptors as X={x _k ,k=1,2,...,K};

(4-2) Use a 50-element mixed Gaussian model u _λ (x) to simulate the generation process of X,

Among them, the parameter λ={ω _l ,μ _l ,σ _l ,l=1,…,50}, ω _l , μ _l and σ _l represent the mixture weight, mean vector and standard deviation of the Gaussian function u _l respectively; the parameter λ is estimated on local feature descriptors generated from a large number of training sets according to the EM criterion;

(4-3) Calculate the partial derivative of the Gaussian model u _λ (x) for μ _l and σ _l respectively, and obtain two gradient vectors:

Among them, γ _k (l) represents the probability that the feature x _k is generated by the l-th Gaussian function;

(4-4) Merge all 50 Gaussian function corresponding gradient vectors to get Fisher Vector code

(4-5) For the above Each dimension m of is subjected to power regularization as follows, where the power index is 0.5:

(4-5) After power regularization conduct Regularization processing to obtain descriptors of blurred image features

Step 5, according to the training data set and the corresponding feature descriptor, adopt a one-to-one multiclass classification strategy to train a linear support vector machine classifier; in this embodiment, open source LIBSVM is adopted to train and establish a support vector machine classifier; The output result of the classifier is the final recognition result.

It is easy for those skilled in the art to understand that the above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, All should be included within the protection scope of the present invention.

Claims (5)

1. A method for obtaining a feature descriptor of a fuzzy invariant image, comprising the following steps: (1) Blur the samples in the source image, and grayscale the blurred image; (2) Using a sliding window to densely sample the image obtained by grayscale processing to obtain a plurality of local image blocks, and divide each local image block into a plurality of unit blocks; (3) extracting LPQ+ for each unit block, and then merging the extracted multiple groups of LPQ+ to obtain the feature descriptor of the local block; (4) Perform Fisher Vector encoding, power regularization and Regularization processing to obtain feature descriptors of blurred images. 2. the method for obtaining the feature descriptor of fuzzy invariant image as claimed in claim 1, is characterized in that, also comprises the step of image recognition after described step (4): (5) According to the training data set and the corresponding feature descriptor, a one-to-one multi-class classification strategy is used to train a linear support vector machine classifier; the output result of the classifier is the image recognition result. 3. the method for obtaining the feature descriptor of fuzzy invariant image as claimed in claim 1 or 2, is characterized in that, described step (2) comprises following substep: (2-1) Using a fixed-size, step-size sliding window to densely sample the image obtained by grayscale processing to obtain multiple local image blocks; (2-2) Divide each local image block into a plurality of small units for storing the spatial structure information of the local block. 4. the method for obtaining the feature descriptor of fuzzy invariant image as claimed in claim 1 or 2, is characterized in that, described step (3) comprises following substep: (3-1) Using STFT to extract the phase information of each pixel in the unit block with the M×M neighborhood around the central pixel as the calculation range: $\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}}{<span\; class="notranslate">\; \&Sigma;</span>}\underset{\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}}{<span\; class="notranslate">\; \&Sigma;</span>}\mathrm{<span\; class="notranslate">\; g</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{\frac{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; j</span>}\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; v</span>}\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; m</span>}}}$ Among them, (x, y), (u, v) refer to the coordinates in the blurred image and the corresponding Fourier transform domain; g(x, y) refers to the blurred image, N _x and N _y refer to the pixel point (x, y) the neighborhood range; According to the four low-frequency points u ₁ =(a,0), u ₂ =(0,a), u ₃ =(a,a), u ₄ =(a,-a), four phase values ∠G( u ₁ ), ∠G(u ₂ ), ∠G(u ₃ ), ∠G(u ₄ ); where a is a constant; (3-2) Press i=1,2,...,I, divide the phase value to obtain I different angle areas, classify the above four phase values into the corresponding angle areas and quantify according to the following voting values, ${\mathrm{<span\; class="notranslate">\; v</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; \&angle;</span>\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}$ Among them, α is an adjustable parameter, Refers to the voting value obtained by ∠G(u _j ) under the corresponding angle area i; where, j=1,2,3,4; (3-3) Obtain four histograms according to the quantized values of the four phase values and merge the histograms to obtain LPQ+; (3-4) Merge the LPQ+ of each unit block of the same local image block to obtain the feature descriptor of the local block. 5. the method for obtaining the feature descriptor of fuzzy invariant image as claimed in claim 1 or 2, is characterized in that, described step (4) comprises following substep: (4-1) Denote the extracted K local block feature descriptors as X={x _k ,k=1,2,...,K}; (4-2) Use N-ary mixed Gaussian model u _λ (x) to simulate the generation process of X, which is denoted as ${\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; l</span>}}{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; l</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span>$ Among them, the parameter λ={ω _l ,μ _l ,σ _l ,l=1,…,N}, ω _l , μ _l and σ _l refer to the mixing weight, mean vector and standard deviation of the Gaussian function u _l respectively; the parameter λ is estimated on local feature descriptors generated from a large training set according to the EM criterion; (4-3) Calculate partial derivatives of Gaussian model u _λ (x) for μ _l and σ _l respectively, and obtain two gradient vectors: Among them, γ _k (l) represents the probability that the feature x _k is generated by the l-th Gaussian function; (4-4) Merge all N Gaussian function corresponding gradient vectors to obtain Fisher Vector code (4-5) right Each dimension m of is subjected to power regularization as follows: (4-6) After the power regularization process conduct Regularization processing, get the descriptor of fuzzy image features: