A clarity evaluation method for light field saliency detection dataset

Abstract

The present invention discloses a method for evaluating the clarity of a light field saliency detection data set. First, three masks are made through a saliency region mask, and the corresponding window function is obtained; the saliency region image of a light field focal stack image is obtained through mask A; the saliency region image of the focal stack is converted from the spatial domain to the frequency domain by Fourier transform, and its spectrum information is obtained; the high-frequency information of the saliency region is inversely Fourier transformed to obtain a spatial domain image; the high-frequency part of the spatial domain is intercepted by an eroded mask B, and the sum of the grayscale values ​​of the high-frequency image is calculated to represent the magnitude of the high-frequency energy value of the saliency region of the focal stack image; the high-frequency energy value of the saliency region of the stack image is normalized to represent the training weight of the image. The present invention determines the clarity of the saliency region in the focal stack of the data set, and uses this as an evaluation standard for the training weight of the image in the focal stack. The clearer the saliency region, the higher the training weight of the image.

Description

Technical Field

[0001] The invention relates to the field of saliency detection, and is a method for evaluating the quality of images of different layers in a focal stack of a light field saliency detection data set. Technical Background

[0002] Saliency monitoring requires corresponding data sets for network training. From the perspective of data structure, there are color data sets, RGB-D data sets, light field saliency detection data sets, etc. In the light field saliency detection data set, each scene exists in the form of a focal stack, and the focal stack in the data set is usually disordered. Therefore, when inputting the network for training, the determination of the weight of each picture in the focal stack is particularly important for network training. Summary of the invention

[0003] In view of the deficiencies in the prior art, the present invention proposes a method for evaluating the clarity of a light field saliency detection dataset.

[0004] Each scene in the light field saliency detection dataset consists of a set of focus stacks, a full focus map, and a salient area mask. Taking a set of focus stacks including N pictures with different focal lengths as an example, a scene is represented by N+2 picture records.

[0005] A method for evaluating the clarity of a light field saliency detection dataset is proposed. The clarity of a salient area in N focal stack images is evaluated to obtain a training weight for each focal stack image. The specific steps are as follows:

[0006] Step 1: Detect salient regions in the dataset using light field saliency mask I mask Make three masks and get the corresponding window function mask circ .

[0007] A scene in the light field saliency detection dataset consists of a set of N images, a focal stack I1, I2, ..., I N , a fully focused image I focus , a salient region mask I mask A total of N+2 images. Mask the salient region I mask Binarization is performed to obtain A mask , that is, A mask =I mask &I mask ;

[0008] A mask Perform an erosion followed by an expansion operation to remove I mask Small and medium connected domains, mask A1 is obtained;

[0009] Dilate A1 to obtain mask A, with a dilation intensity of 3 to 5 pixels;

[0010] Perform an erosion operation on A1 to obtain mask B, with an erosion intensity of 3 to 5 pixels;

[0011] Window function mask circ and salient region mask I mask The same size, and the window is located in the mask circ central;

[0012] Step 2: Obtain the salient area image of the light field focus stack image through mask A.

[0013] The salient area of ​​the focus stack image and the surrounding part of the image are extracted separately through mask A to obtain the focus stack salient area image in And i=1, 2,...,N;

[0014] Step 3: Use Fourier transform to transform the focal stack salient area image Convert from spatial domain to frequency domain to obtain its spectrum information.

[0015] Step 4: Perform inverse Fourier transform on the high-frequency information of the salient area to obtain the spatial domain image in And i=1, 2,...,N;

[0016] Step 5: Use the eroded mask B to intercept the high-frequency part of the airspace and obtain the high-frequency image after the significant area is intercepted in And i = 1,, 2, ..., N. Calculate The sum of gray values ​​m1, m2, ..., m N , which is used to represent the high-frequency energy value of the salient area of ​​the focal stack image.

[0017] Step 6: Normalize the high-frequency energy values ​​of the salient regions of the stacked images to represent the training weights M1, M2, ..., M of the image. N ,in

[0018] Furthermore, the specific method of step three is as follows:

[0019] For the data obtained in step 2 Perform a two-dimensional Fourier transform to obtain the spectrum in is the Fourier operator. Use the window function mask circ Focus stack salient region spectrum information Perform high-pass filtering to obtain high-frequency information of significant areas in And i=1, 2,...,N.

[0020] The beneficial effects of the present invention are as follows:

[0021] The present invention determines the clarity of a significant area in a focal stack of a data set, and uses this as an evaluation criterion for the training weight of the image in the focal stack. The clearer the significant area, the higher the training weight of the image. DETAILED DESCRIPTION

[0022] A method for evaluating the clarity of a light field saliency detection dataset is provided. The method evaluates the clarity of a salient area in N focal stack images and obtains a training weight for each focal stack image. The specific steps are as follows:

[0023] Step 1: Detect salient regions in the dataset using light field saliency mask I mask Make three masks and get the corresponding window function mask circ .

[0024] A scene in the light field saliency detection dataset consists of a set of N images, each of which is a focal stack I1, I2, ..., I N , a fully focused image I focus , a salient region mask I mask A total of N+2 images. Mask the salient region I mask Binarization is performed to obtain A mask , that is, A mask =I mask &I mask ;

[0025] A mask Perform an erosion followed by an expansion operation to remove I mask Small and medium connected domains, mask A1 is obtained;

[0026] Dilate A1 to obtain mask A, with a dilation intensity of 3 to 5 pixels;

[0027] Perform an erosion operation on A1 to obtain mask B, with an erosion intensity of 3 to 5 pixels;

[0028] Window function mask circ and salient region mask I mask The same size, and the window is located in the mask circ central;

[0029] Step 2: Obtain the salient area image of the light field focus stack image through mask A.

[0030] The salient area of ​​the focus stack image and the surrounding part of the image are extracted separately through mask A to obtain the focus stack salient area image in And i=1, 2,...,N;

[0031] Step 3: Use Fourier transform to transform the focal stack salient area image Convert from spatial domain to frequency domain to obtain its spectrum information.

[0032] For the data obtained in step 2 Perform a two-dimensional Fourier transform to obtain the spectrum in is the Fourier operator. Use the window function mask circ Focus stack salient region spectrum information Perform high-pass filtering to obtain high-frequency information of significant areas in And i=1, 2,...,N.

[0033] Step 4: Perform inverse Fourier transform on the high-frequency information of the salient area to obtain the spatial domain image in And i=1, 2,...,N;

[0034] Step 5: Use the eroded mask B to intercept the high-frequency part of the airspace and obtain the high-frequency image after the significant area is intercepted in And i = 1,, 2, ..., N. Calculate The sum of gray values ​​m1, m2, ..., m N , which is used to represent the high-frequency energy value of the salient area of ​​the focal stack image.

[0035] Step 6: Normalize the high-frequency energy values ​​of the salient regions of the stacked images to represent the training weights M1, M2, ..., M of the image. N ,in

Claims

1. A method for evaluating the clarity of a light field saliency detection dataset, characterized in that: Evaluate the clarity of the salient areas in N focal stack images and obtain the training weight of each focal stack image. The specific steps are as follows: Step 1: Detect salient regions in the dataset using light field saliency mask I mask Make three masks and get the corresponding window function mask circ ; A scene in the light field saliency detection dataset consists of a set of N images, each of which is a focal stack I1, I2, ..., I N , a fully focused image I focus , a salient region mask I mask A total of N+2 images; the salient area mask I mask Binarization is performed to obtain A mask , that is, A mask =I mask &I mask ; A mask Perform an erosion followed by an expansion operation to remove I mask Small and medium connected domains, mask A1 is obtained; Dilate A1 to obtain mask A, with a dilation intensity of 3 to 5 pixels; Perform an erosion operation on A1 to obtain mask B, with an erosion intensity of 3 to 5 pixels; Window function mask circ and salient region mask I mask The same size, and the window is located in the mask circ central; Step 2: Obtain a salient area image of the light field focus stack image through mask A; The salient area of ​​the focus stack image and the surrounding part of the image are extracted separately through mask A to obtain the focus stack salient area image in And i=1, 2,...,N; Step 3: Use Fourier transform to transform the focal stack salient area image Convert from the spatial domain to the frequency domain to obtain its spectrum information, where i = 1, 2, ..., N; Step 4: Perform inverse Fourier transform on the high-frequency information of the salient area to obtain the spatial domain image in And i=1, 2,...,N; Step 5: Use the eroded mask B to intercept the high-frequency part of the airspace and obtain the high-frequency image after the significant area is intercepted in and i = 1,, 2, ..., N; calculate The sum of gray values ​​m1, m2, ..., m N , used to represent the high-frequency energy value of the salient area of ​​the focal stack image; Step 6: Normalize the high-frequency energy values ​​of the salient regions of the stacked images to represent the training weights M1, M2, ..., M of the image. N ,in 2. The method for evaluating clarity of a light field saliency detection dataset according to claim 1, characterized in that: The specific method of step three is as follows: For the data obtained in step 2 Perform a two-dimensional Fourier transform to obtain the spectrum in is the Fourier operator; use the window function mask circ Focus stack salient region spectrum information Perform high-pass filtering to obtain high-frequency information of significant areas in And i=1, 2,...,N.

Images