Depth information importance evaluation method inspired by brain mechanism

Abstract

The invention discloses a depth information importance evaluation method inspired by a brain mechanism, and the method employs filters obtained in two modes to filter a color input image, select information related to a current task, ignore information irrelevant to the task, retain important information, and suppress unimportant information. Therefore, a foundation is laid for sorting and selection of information importance. The obtained feature map subjected to importance evaluation contains richer core information, background noise can be removed to the maximum extent, a core target object is reserved, a solid foundation is laid for further application to target recognition and image retrieval, the obtained target object is more complete and clearer, and the background noise is smaller.

Description

Technical Field

[0001] The present invention relates to the field of image retrieval technology, and in particular to a method for evaluating the importance of depth information inspired by brain mechanisms. Background Art

[0002] Vision is the basis of the brain's advanced cognitive functions. The study of visual neural mechanisms has become a hot topic in the fields of pattern recognition, computer vision, cognitive science, and visual neuroscience. There is a deep relationship between deep neural networks and human thinking, and deep neural networks are developed under the inspiration of brain neuroscience. How to establish a visual computing model of brain-like intelligence is a topic that is worthy of in-depth research. The research findings on brain mechanisms provide new theoretical basis and guidance for the establishment of newer visual computing models and deep learning models, and visual computing models and deep learning models can guide us to simulate brain mechanisms and propose new solutions and new research ideas.

[0003] The human brain's information processing and cognitive functions are extremely complex. Judging the importance of information is a high-level brain function that helps people better adapt to complex environments and controls their attention and learning abilities. By studying how the brain dynamically evaluates the importance of external information and the intrinsic mechanism of biological significance, and establishing a visual computing model, it is of great significance to narrow the semantic gap and improve object recognition and image retrieval performance. Summary of the invention

[0004] The present invention aims to solve the problem that there is a large semantic gap between the feature vectors used in image retrieval and the human brain's understanding of images, which affects the image retrieval performance, and provides a method for evaluating the importance of deep information inspired by the brain mechanism.

[0005] To solve the above problems, the present invention is achieved through the following technical solutions:

[0006] A method for evaluating the importance of deep information inspired by brain mechanisms includes the following steps:

[0007] Step 1: Segment the color input image X into color input images including N superpixel regions using the SLIC superpixel segmentation method;

[0008] Step 2: Calculate the saliency map S of each superpixel region i , and then the saliency map S of N superpixel regions i After integration, the saliency map S of the color input image is obtained;

[0009] Step 3: Send the color input image X into the ViT model to obtain the visual conversion feature map And combine the saliency map S with the visual conversion feature map Convolution is performed to obtain a salient feature map

[0010] Step 4: Analyze the significant feature map Perform dimensionality reduction and normalization to obtain the first filter f 1 ; and the first filter f 1 Convolve with the color input image X to obtain the first filter feature map X1;

[0011] Step 5: Analyze the significant feature map Perform feature extraction and normalization to obtain the second filter f 2 ; and the second filter f 2 Convolve with the color input image X to obtain the second filtered feature map X2;

[0012] Step 6: Superimpose the first filtering feature map X1 and the second filtering feature map X2 to obtain an importance feature map Ω, and use the importance feature map Ω as an evaluation result of the color input image;

[0013] In the above, i=1, 2, ..., N, where N is the number of set superpixel regions.

[0014] In step 2, the superpixel region The salient map S i for:

[0015]

[0016] In the formula, Represents superpixel area and The difference between Lab color space, Represents superpixel area and The difference in HSV color space; Represents superpixel area and The best direction of Represents superpixel area The average value of the horizontal coordinates of all pixels in the color input image X. Represents superpixel area The average value of the ordinates corresponding to all pixels in the color input image X; Represents superpixel area The average value of the horizontal coordinates of all pixels in the color input image X. Represents superpixel area The average value of the ordinates of all pixels in the color input image X; Δε is the set spatial displacement threshold;

[0017] i,j=1,2,…,N,j≠i, N is the number of set superpixel areas.

[0018] Superpixel area and The best direction The determination process is as follows:

[0019] Step 1) inputting a color input image including N superpixel regions into a directional feature map extraction submodel of the Itti visual saliency model to obtain M directional feature maps, each of which includes N superpixel regions;

[0020] Step 2) calculating the directional feature value of each superpixel region in each directional feature map, wherein the directional feature value is the sum of the pixel values ​​of all pixels in the current superpixel region of the current directional feature map;

[0021] Step 3) For each superpixel region, find the directional feature map where the maximum directional feature value is located, and use the direction of the directional feature map as the main direction of the superpixel region;

[0022] Step 4) Compare the main directions of the two superpixel regions: If the main directions of the two superpixel regions are the same, then the main direction is the optimal direction of the two superpixel regions. If the main directions of two superpixel regions are different, the average angle of the main directions of the two superpixel regions is the optimal direction of the two superpixel regions.

[0023] In the above, i, j = 1, 2, ..., N, j ≠ i, N is the number of set superpixel regions, and M is the number of set directional feature maps.

[0024] Superpixel area and Lab color space difference for:

[0025]

[0026] In the formula, L, a, b represent the three color channels of Lab color. Represents superpixel area The color value of the kth color channel, Represents superpixel area The color value of the kth color channel; Represents superpixel area The excitatory response, Represents superpixel area excitatory response; Represents superpixel area The inhibitory response Represents superpixel area inhibitory response; represents the Kronecker product, ‖*‖ 2 represents the Euclidean distance; i, j = 1, 2, ..., N, j ≠ i, N is the number of superpixel regions set.

[0027] Superpixel area and The difference between HSV color space for:

[0028]

[0029] In the formula, h, s, v represent the three color channels of HSV color. Represents superpixel area The color value of the kth color channel, Represents superpixel area The color value of the kth color channel; Represents superpixel area The excitatory response, Represents superpixel area excitatory response; Represents superpixel area The inhibitory response Represents superpixel area inhibitory response; represents the Kronecker product, ‖*‖ 2 represents the Euclidean distance; i, j = 1, 2, ..., N, j ≠ i, N is the number of set superpixel regions; i, j = 1, 2, ..., N, j ≠ i, N is the number of set superpixel regions.

[0030] Superpixel area Excitatory response for:

[0031]

[0032] Superpixel area The inhibitory response for:

[0033]

[0034] In the formula, Represents superpixel area The horizontal coordinate of a pixel point in the color input image X, Represents superpixel area The ordinate corresponding to a certain pixel point in the color input image X; Represents superpixel area The average value of the horizontal coordinates of all pixels in the color input image X. Represents superpixel area The average value of the ordinates of all pixels in the color input image X; i Super pixel area The standard deviation of

[0035] ε is the set adjustment coefficient, Represents superpixel area The length of the major axis, Represents superpixel area The short axis length of ; i = 1, 2, ..., N, N is the number of super pixel areas set.

[0036] Superpixel area Excitatory response for:

[0037]

[0038] Superpixel area The inhibitory response for:

[0039]

[0040] In the formula, Represents superpixel area The horizontal coordinate of a pixel point in the color input image X, Represents superpixel area The ordinate corresponding to a certain pixel point in the color input image X; Represents superpixel area The average value of the horizontal coordinates of all pixels in the color input image X. Represents superpixel area The average value of the ordinates of all pixels in the color input image X; j Super pixel area The standard deviation of ε is the set adjustment coefficient, Represents superpixel area The length of the major axis, Represents superpixel area The short axis length of ; j = 1, 2, ..., N, N is the number of super pixel areas set.

[0041] Compared with the prior art, the feature map obtained by the present invention after importance evaluation contains richer core information, can remove background noise to the greatest extent, and retain the core target object, laying a solid foundation for further application in target recognition and image retrieval. The target object obtained is more complete and clearer, and the background noise is smaller. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Schematic diagram of a method for evaluating the importance of deep information inspired by brain mechanisms. DETAILED DESCRIPTION

[0043] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in combination with specific examples and with reference to the accompanying drawings.

[0044] In the primary visual cortex, visual neurons selectively respond to color information and shape information, they have direction selectivity, and there is overlap between the direction pathway and the color pathway. Due to the complex relationship and interaction between color tuning, direction selection and cytochrome oxidase histology, they enable individual neurons to accurately and clearly encode shape and color. In order to simulate the above neuronal mechanism, the present invention proposes a method for evaluating the importance of depth information inspired by the brain mechanism, such as Figure 1 As shown, the steps include:

[0045] Step 1: Use the SLIC superpixel segmentation method to segment the color input image X into a color input image including N superpixel regions, where i = 1, 2, ..., N, and N is the number of superpixel regions, which is a set value determined by experiments.

[0046] SLIC (simple linear iterative clustering) is a superpixel segmentation algorithm based on K-means clustering proposed in 2010. It is simple in concept and easy to implement. It converts color images into 5-dimensional feature vectors in CIELAB color space and XY coordinates, and constructs a distance metric for the 5-dimensional feature vectors to perform local clustering of image pixels. The SLIC superpixel segmentation method can generate compact, approximately uniform superpixels, and has a high comprehensive evaluation in terms of computing speed, object contour preservation, and superpixel shape, which is more in line with people's expected segmentation effect. In most cases, superpixels are some irregular image blocks or areas, in which represents the major axis length of the superpixel area, Represents the minor axis length of the superpixel area.

[0047] Step 2: Calculate the saliency map S of each superpixel region i , and then the saliency map S of N superpixel regions i After integration, the saliency map S of the color input image is obtained.

[0048] Because there are often multiple objects in an image or scene, the brain must first evaluate the importance of the relevant information (multiple objects) before making a selection. Studies have shown that the importance of external stimuli (i.e., biological significance) depends not only on the physical characteristics of the stimulus itself, but also on the animal's internal physiological state and the external environment in which it is located. Therefore, the present invention applies the saliency map S to the information importance assessment, and the object saliency detection method proposed is not only for finding the so-called foreground or object, but for realizing the ranking of the importance of multiple objects based on the foreground or object. Therefore, the present invention does not directly draw on deep learning technology to perform saliency detection, which is obviously different from the saliency detection method based on deep learning or traditional manual features.

[0049] Considering that there may be multiple shape regions or objects in an image or scene, the present invention uses the spatial arrangement and color difference of the shape regions or objects to jointly determine the saliency, that is, uses the directional feature saliency map and the two color differences to determine the object saliency.

[0050] The above super pixel area The salient map S i for:

[0051]

[0052] In the formula, Represents superpixel area and The difference between Lab color space, Represents superpixel area and The difference in HSV color space; Represents superpixel area and The best direction of max{d(R i ,R j ,θ),Δε} is used to constrain the superpixel region R i and R j The spatial displacement (distance) between them is no more than Δε, and the direction of the superpixel region where there is spatial displacement is defined as θ, and Δε is the set spatial displacement threshold, whose value is determined by experiments.

[0053]

[0054] In the formula, Represents superpixel area The average value of the horizontal coordinates of all pixels in the color input image X. Represents superpixel area The average value of the ordinates corresponding to all pixels in the color input image X; Represents superpixel area The average value of the horizontal coordinates of all pixels in the color input image X. Represents superpixel area The average value of the vertical coordinates of all pixels in the color input image X; i, j = 1, 2, ..., N, j ≠ i, N is the number of super pixel areas set.

[0055] (1) Superpixel area and The best direction

[0056] Superpixel area and The best direction The determination process is as follows:

[0057] 1) It will include N superpixel regions The color input image is input into the directional feature map extraction submodel of the Itti visual saliency model to obtain M directional feature maps S(θ v ), each directional feature map S(θ v ) include N superpixel regions Wherein v = 1, 2, ..., M, M is the number of directional characteristic maps, which is a set value determined by experiments.

[0058] The Itti visual saliency model is a visual attention model designed by Itti et al. in 1998 based on the visual nervous system of early primates. The model first uses the Gaussian sampling method to construct a Gaussian pyramid of the color, brightness and direction of the image, and then uses the Gaussian pyramid to calculate the brightness feature map, color feature map and direction feature map of the image. Finally, the brightness, color and direction feature maps are obtained by combining feature maps of different scales, and the final visual saliency map is obtained by adding them together. However, the present invention does not need to obtain the final visual saliency map, only the direction feature map, so the present invention does not use the complete Itti visual saliency model to obtain the visual saliency map of the image, but only uses the direction feature map extraction submodel used to extract the direction feature map in the Itti visual saliency model to obtain the direction feature map of the image.

[0059] To simulate the directional selectivity of visual cortical neurons, the set direction can be represented by θ v ∈{θ 1 ,θ 2,…,θ M In this embodiment, M=4 is set, and the direction set at this time can be expressed as The directional feature map can be expressed as S(0), Since the color input image X input to the directional feature map extraction sub-model contains N superpixel regions Therefore, each directional feature map S(θ v ) also includes N superpixel regions

[0060] 2) Calculate each directional feature map S(θ v ) in each superpixel region The directional feature value of the vth directional feature map S(θ v )’s i-th superpixel region The directional feature value is the vth directional feature map S(θ v )’s i-th superpixel region The sum of the pixel values ​​of all pixels.

[0061] 3) For each superpixel region Find the directional feature map with the maximum directional eigenvalue, and use the direction of the directional feature map as the superpixel area The main direction O i (θ).

[0062] 4) Compare 2 superpixel regions and The main direction O i (θ) and O j (θ):

[0063] If the superpixel area The main direction O i (θ) and superpixel area The main direction O j (θ)same(O i (θ) = O j (θ)), then the optimal direction

[0064] If the superpixel area The main direction O i (θ) and superpixel area The main direction O j (θ) different (O i (θ)≠O j (θ)), then the optimal direction

[0065] Since there is a very close relationship between spatial connectivity, directional selectivity and shape features, in order to match the spatial connectivity with the directional selectivity of the visual cortex, the present invention determines the optimal direction of the superpixel region with existing spatial displacement through a directional feature saliency map.

[0066] (2) Superpixel area and Lab color space difference

[0067] In Lab color space, the superpixel area and Lab color space difference for:

[0068]

[0069] In the formula, L, a, b represent the three color channels of Lab color. Represents superpixel area The color value of the kth color channel, Represents superpixel area The color value of the kth color channel; Represents superpixel area The excitatory response, Represents superpixel area excitatory response; Represents superpixel area The inhibitory response Represents superpixel area inhibitory response; represents the Kronecker product, ‖*‖ 2 represents the Euclidean distance; i, j = 1, 2, ..., N, j ≠ i, N is the number of superpixel regions set.

[0070] (3) Superpixel area and The difference between HSV color space

[0071] In HSV color space, superpixel area and The difference between HSV color space for:

[0072]

[0073] In the formula, h, s, v represent the three color channels of HSV color. Represents superpixel area The color value of the kth color channel, Represents superpixel area The color value of the kth color channel; Represents superpixel area The excitatory response, Represents superpixel area excitatory response; Represents superpixel area The inhibitory response Represents superpixel area inhibitory response; represents the Kronecker product, ‖*‖ 2 represents the Euclidean distance; i, j = 1, 2, ..., N, j ≠ i, N is the number of set superpixel regions; i, j = 1, 2, ..., N, j ≠ i, N is the number of set superpixel regions.

[0074] (4) Excitatory response and inhibitory responses

[0075] Since the directional selectivity of the postsynaptic neuron is independent of the selectivity of the presynaptic neuron, but is related to the spatial displacement between the excitatory and inhibitory presynaptic neurons, the present invention introduces excitatory and inhibitory responses to characterize the existence of a certain spatial displacement between different superpixels. The excitatory response and the inhibitory response are two Gaussian filters with different structures.

[0076] Superpixel area Excitatory response for:

[0077]

[0078] Superpixel area The inhibitory response for:

[0079]

[0080] In the formula, Represents superpixel area The horizontal coordinate of a pixel point in the color input image X, Represents superpixel area The ordinate corresponding to a certain pixel point in the color input image X; Represents superpixel area The average value of the horizontal coordinates of all pixels in the color input image X. Represents superpixel area The average value of the ordinates of all pixels in the color input image X; δ is the standard deviation, ε is the set adjustment coefficient, Represents superpixel area The length of the major axis, Represents superpixel area The short axis length of ; i = 1, 2, ..., N, N is the number of super pixel areas set.

[0081] Superpixel area Excitatory response for:

[0082]

[0083] Superpixel area The inhibitory response for:

[0084]

[0085] In the formula, Represents superpixel area The horizontal coordinate of a pixel point in the color input image X, Represents superpixel area The ordinate corresponding to a certain pixel point in the color input image X; Represents superpixel area The average value of the horizontal coordinates of all pixels in the color input image N. Represents superpixel area The average value of the ordinates of all pixels in the color input image X; j Super pixel area The standard deviation of

[0086] ε is the set adjustment coefficient, Represents superpixel area The length of the major axis, Represents superpixel area The short axis length of ; j = 1, 2, ..., N, N is the number of super pixel areas set.

[0087] Step 3: Send the color input image X into the ViT model to obtain the visual conversion feature map And combine the saliency map S with the visual conversion feature map Convolution is performed to obtain a salient feature map

[0088] The ViT (Vision Transformer) model was proposed by the Google team at ICLR 2021. It is mainly used to extract image features. Its idea is simple, effective and has good scalability. It is regarded as one of the important milestones in the application of Transformer in the field of computer vision. The present invention uses the ViT model to process the color input image X, where the ViT preprocessing model uses the ViT-Large model, and the data output by its last layer is defined as the visual transformation feature map Where W and H are the width and height of the feature map respectively, K is the number of feature maps, K=1024.

[0089] The saliency map S of the color input image is used as a filter and the visual conversion feature map Before convolution, the saliency map S of the color input image needs to be scaled so that its width and height are consistent with the visual conversion feature map The width and height of Will As a filter with The process of convolution is called saliency filtering, and its result is called saliency feature map (shape feature map)

[0090]

[0091] The present invention utilizes the feature map output by the ViT model, constructs a filter based on a saliency map by simulating the direction selectivity of visual cortical neurons, utilizes the saliency map to select important information, and extracts shape features (shape regions or objects).

[0092] Step 4: Analyze the significant feature map Perform dimensionality reduction and normalization to obtain the first filter f 1 ; and the first filter f 1 Convolve with the color input image X to obtain the first filtered feature map X1.

[0093] The present invention uses principal component analysis (PCA) to analyze the significant feature map Dimensionality reduction is performed by retaining a certain proportion of principal components to obtain the mapping image of each significant feature map. The optimal information retention amount of the PCA algorithm needs to be determined through experiments. In this embodiment, 20% of the principal components are intended to be retained. Perform dimensionality reduction to obtain 1024 mapping images Then overlay all 1024 mapping images X 1 A superimposed image can be obtained And for image P 1 Normalization can be performed to obtain the filter

[0094] Step 5: Analyze the significant feature map Perform feature extraction and normalization to obtain the second filter f 2 ; and the second filter f 2 Convolve with the color input image X to obtain the second filtered feature map X2.

[0095] The present invention adopts MAC or SPoC feature extraction method to extract the significant feature map. In this implementation, we first extract features from the salient feature map. Perform feature extraction to obtain 1024 special images Then each Jute image X 2 Add the pixel values ​​of all pixels to get the polygraph image X 2 The total pixel value of the image is calculated, and the jute images are sorted in descending order according to the total pixel value. Then, n jute images with larger total pixel values ​​are selected. 2 By superimposing, a superimposed image can be obtained. And the superimposed image P 2 Normalization can be performed to obtain the filter The value of n is determined through experiments. In this embodiment, n=50.

[0096] Step 6: Superimpose the first filtering feature map X1 and the second filtering feature map X2 to obtain an importance feature map Ω, and use the importance feature map Ω as an evaluation result of the color input image.

[0097] The filter f obtained by the present invention adopts two methods 1 and f 2 To filter the color input image X, select information related to the current task, ignore information irrelevant to the task, retain important information, and suppress unimportant information, thus laying the foundation for the sorting and selection of information importance. The important feature map Ω is a feature map that has been evaluated for information importance. It contains richer core information, can remove background noise to the greatest extent, and retain the core target object, laying a solid foundation for further application in target recognition and image retrieval.

[0098] It should be noted that although the embodiments of the present invention described above are illustrative, they are not intended to limit the present invention, and therefore the present invention is not limited to the above specific embodiments. Without departing from the principles of the present invention, any other embodiments obtained by those skilled in the art under the guidance of the present invention are deemed to be within the protection of the present invention.

Claims

1. A method for evaluating the importance of deep information inspired by brain mechanisms, characterized by: The steps include: Step 1: Segment the color input image X into color input images including N superpixel regions using the SLIC superpixel segmentation method; Step 2: Calculate the saliency map S of each superpixel region i , and then the saliency map S of N superpixel regions i After integration, the saliency map S of the color input image is obtained; Step 3: Send the color input image X into the ViT model to obtain the visual conversion feature map And combine the saliency map S with the visual conversion feature map Convolution is performed to obtain a salient feature map Step 4: Analyze the significant feature map Perform dimensionality reduction and normalization processing to obtain a first filter f1; and convolve the first filter f1 with the color input image X to obtain a first filter feature map X1; Step 5: Analyze the significant feature map Perform feature extraction and normalization processing to obtain a second filter f2; and convolve the second filter f2 with the color input image X to obtain a second filter feature map X2; Step 6: Superimpose the first filtering feature map X1 and the second filtering feature map X2 to obtain an importance feature map Ω, and use the importance feature map Ω as an evaluation result of the color input image; In the above, i=1, 2, ..., N, where N is the number of set superpixel regions.

2. According to claim 1, a method for evaluating the importance of depth information inspired by brain mechanisms is characterized in that: In step 2, the superpixel region The salient map S i for: In the formula, Represents superpixel area and The difference between the Lab color space, Represents superpixel area and The difference in HSV color space; Represents superpixel area and The best direction of Represents superpixel area The average value of the horizontal coordinates of all pixels in the color input image X. Represents superpixel area The average value of the ordinates corresponding to all pixels in the color input image X; Represents superpixel area The average value of the horizontal coordinates of all pixels in the color input image X. Represents superpixel area The average value of the ordinates of all pixels in the color input image X; Δε is the set spatial displacement threshold; i,j=1,2,…,N,j≠i, N is the number of set superpixel areas.

3. The method for evaluating the importance of depth information inspired by brain mechanism according to claim 2, characterized in that: Superpixel area and The best direction The determination process is as follows: Step 1) inputting a color input image including N superpixel regions into a directional feature map extraction submodel of the Itti visual saliency model to obtain M directional feature maps, each of which includes N superpixel regions; Step 2) calculating the directional feature value of each superpixel region in each directional feature map, wherein the directional feature value is the sum of the pixel values ​​of all pixels in the current superpixel region of the current directional feature map; Step 3) For each superpixel region, find the directional feature map where the maximum directional feature value is located, and use the direction of the directional feature map as the main direction of the superpixel region; Step 4) Compare the main directions of the two superpixel regions: If the main directions of the two superpixel regions are the same, then the main direction is the optimal direction of the two superpixel regions. If the main directions of two superpixel regions are different, the average angle of the main directions of the two superpixel regions is the optimal direction of the two superpixel regions. In the above, i, j = 1, 2, ..., N, j ≠ i, N is the number of set superpixel regions, and M is the number of set directional feature maps.

4. The method for evaluating the importance of depth information inspired by brain mechanism according to claim 2, characterized in that: Superpixel area and Lab color space difference for: In the formula, L, a, b represent the three color channels of Lab color. Represents superpixel area The color value of the kth color channel, Represents superpixel area The color value of the kth color channel; Represents superpixel area The excitatory response, Represents superpixel area excitatory response; Represents superpixel area The inhibitory response Represents superpixel area inhibitory response; represents the Kronecker product, ‖*‖ 2 represents the Euclidean distance; i, j = 1, 2, ..., N, j ≠ i, N is the number of superpixel regions set.

5. The method for evaluating the importance of depth information inspired by brain mechanism according to claim 2, characterized in that: Superpixel area and The difference between HSV color space for: In the formula, h, s, v represent the three color channels of HSV color. Represents superpixel area The color value of the kth color channel, Represents superpixel area The color value of the kth color channel; Represents superpixel area The excitatory response, Represents superpixel area excitatory response; Represents superpixel area The inhibitory response Represents superpixel area inhibitory response; represents the Kronecker product, ‖*‖ 2 represents the Euclidean distance; i, j = 1, 2, ..., N, j ≠ i, N is the number of set superpixel regions; i, j = 1, 2, ..., N, j ≠ i, N is the number of set superpixel regions.

6. A method for evaluating the importance of depth information inspired by brain mechanisms according to claim 4 or 5, characterized in that: Superpixel area Excitatory response for: Superpixel area The inhibitory response for: In the formula, Represents superpixel area The horizontal coordinate of a pixel point in the color input image X, Represents superpixel area The ordinate corresponding to a certain pixel point in the color input image X; Represents superpixel area The average value of the horizontal coordinates of all pixels in the color input image X. Represents superpixel area The average value of the ordinates of all pixels in the color input image X; i Super pixel area The standard deviation of ε is the set adjustment coefficient, Represents superpixel area The length of the major axis, Represents superpixel area The short axis length of ; i = 1, 2, ..., N, N is the number of super pixel areas set.

7. A method for evaluating the importance of depth information inspired by brain mechanisms according to claim 4 or 5, characterized in that: Superpixel area Excitatory response for: Superpixel area The inhibitory response for: In the formula, Represents superpixel area The horizontal coordinate of a pixel point in the color input image X, Represents superpixel area The ordinate corresponding to a certain pixel point in the color input image X; Represents superpixel area The average value of the horizontal coordinates of all pixels in the color input image X. Represents superpixel area The average value of the ordinates of all pixels in the color input image X; j Super pixel area The standard deviation of ε is the set adjustment coefficient, Represents superpixel area The length of the major axis, Represents superpixel area The short axis length of ; j = 1, 2, ..., N, N is the number of super pixel areas set.

Images