An image retrieval method that simulates the brain's representation of abstract knowledge

Abstract

The present invention discloses an image retrieval method that simulates the brain's representation of abstract knowledge. The method adopts the SimCLRV2 self-supervised learning framework combined with the SwinTransformer network to construct a feature extraction model, and uses the feature extraction model to extract the feature map and depth feature histogram of the image. Then, the direction histogram is constructed by finding the spatial similarity structure of the feature map to jointly encode abstract information and spatial information. The depth feature histogram and the direction histogram are jointly used as the abstract representation of the image for retrieval, which greatly improves the accuracy of image retrieval.

Description

Technical Field

[0001] The present invention relates to the field of image retrieval technology, and in particular to an image retrieval method that simulates the brain's representation of abstract knowledge. Background Art

[0002] In the field of artificial intelligence, there are significant differences between the visual features extracted by machines and human visual perception and cognitive habits. This is primarily because these visual features do not fully mimic the brain's mechanisms for extracting characteristic image content. Abstract representations in the brain naturally emerge from learning, facilitating multiple classifications of the same stimulus. Human hippocampal neurons are capable of generating abstract representations during reasoning. Research has shown that hippocampal neurons can jointly encode abstract and spatial information, with their internal representations exhibiting geometric structures and strong interdependencies. Hippocampal neurons are capable of representing spatial positions and encoding abstract cognitive variables, possessing the ability to simultaneously encode both objective external information and subjective mental intentions. Simulating the brain to represent abstract knowledge is a challenging problem with significant scientific value and research significance.

[0003] In traditional supervised learning, a large amount of labeled data is required to guide model training. However, in practical applications, obtaining large-scale labeled data is very difficult and requires a lot of computing resources. In 2020, Turing Award winner and father of deep learning Geoffrey Hinton and his team proposed a self-supervised learning framework (SimCLRV2), which learns visual representations without human supervision. However, how to learn image representations by drawing on the SimCLRV2 framework, simulate the way the brain represents abstract knowledge, jointly encode spatial information and depth information, and ultimately form abstract feature descriptors and apply them to image retrieval to improve retrieval accuracy remains a difficult problem. Summary of the Invention

[0004] The present invention aims to solve the problem that existing image retrieval methods cannot jointly encode spatial information and depth information, and to provide an image retrieval method that simulates the brain's representation of abstract knowledge.

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

[0006] An image retrieval method that simulates the brain's representation of abstract knowledge includes the following steps:

[0007] Step 1: Use the training image set to train the SimCLRV2 self-supervised learning framework to obtain a trained SimCLRV2 self-supervised learning framework;

[0008] Step 2: Each retrieval image in the retrieval image database is fed into the trained SimCLRV2 self-supervised learning framework, which outputs a feature map set and a deep feature histogram of the retrieval image.

[0009] Step 3: For each feature map in the feature map set of each search image, perform Gabor filtering on it using a two-dimensional Gabor filter to obtain multiple reference directional feature maps of the feature map;

[0010] Step 4: For each feature map, the pixel values ​​of the corresponding pixels of all reference directional feature maps are accumulated and used as the values ​​of the corresponding pixels of the directional feature map to obtain the directional feature map of the feature map, and then obtain the directional feature map set of each search image;

[0011] Step 5: For each feature map, the direction corresponding to the pixel maximum value of the corresponding pixel points of all reference directional feature maps is used as the assignment of the corresponding pixel point of the directional map to obtain the directional map of the feature map, and then obtain the directional map set of each search image;

[0012] Step 6: For each feature map in the feature map set of each search image, record the direction of the local structure with a sense of direction of each pixel point; at the same time, for each directional feature map in the directional feature map set of each search image, record the direction of the local structure with a sense of direction of each pixel point;

[0013] Step 7: Count the number of occurrences of the directions of the spatially similar structures of each retrieval image, and construct a direction histogram using the number of occurrences of the directions of the spatially similar structures recorded for each retrieval image:

[0014] Step 8: Concatenate the depth feature histogram of each retrieval image obtained in step 2 and the direction histogram of each retrieval image obtained in step 7 to obtain a multi-tuple feature of the retrieval image; and normalize the multi-tuple feature of the retrieval image to obtain an image abstract representation of the retrieval image.

[0015] Step 9: Perform steps 2 to 8 on the image to be retrieved to obtain an image abstract representation of the image to be retrieved, calculate the L2 distance between the image abstract representation of the image to be retrieved and the image abstract representations of each retrieval image in the retrieval image database, and output the retrieval image with the smallest L2 distance as the retrieval result.

[0016] In step 1 above, the basic encoder of the SimCLRV2 self-supervised learning framework is the SwinTransformer network.

[0017] In the above step 1, when training the SimCLRV2 self-supervised learning framework, in addition to using the original image expansion method to expand the training image set, the superpixel segmentation method is also used to expand the training image set.

[0018] In the above step 3, after Gabor filtering, each feature map obtains four reference direction feature maps, namely, a 0° reference direction feature map, a 45° reference direction feature map, a 90° reference direction feature map, and a 135° reference direction feature map.

[0019] In the above step 6, for a certain coordinate in the feature map or directional feature map (x o ,y o ) has a pixel value of p0; the pixel with the coordinate (x o ,y o ) of the pixel point, the two adjacent pixels have pixel values ​​of p o+1 and p o-1 ; If the pixel values ​​of these three pixels are equal, that is, p o-1 =p o =p o+1 , then it is considered that these three pixels form a local structure with a sense of direction, and the direction of the local structure with a sense of direction is θ. The direction θ of the local structure with a sense of direction is 0°, 45°, 90° or 135°.

[0020] In the above step 7, the process of counting the number of occurrences of the direction of the spatially similar structure of each retrieval image is as follows: correspondingly traverse each pixel point of each feature map in the feature map set of the retrieval image and each pixel point of each directional feature map in the directional feature map set of the retrieval image: if the direction of the local structure with a sense of direction of the current pixel point of the current feature map of the feature map set is the same as the direction of the local structure with a sense of direction of the current pixel point of the current directional feature map of the directional feature map set, and the direction of the local structure with a sense of direction is the same as the direction of the corresponding pixel point of the corresponding directional mapping map of the directional feature map set, then the two local structures with a sense of direction are considered to be spatially similar structures, and the direction of the local structure with a sense of direction is used as the direction of the spatially similar structure, and the number of occurrences of the direction of the spatially similar structure is accumulated by 1; after the traversal is completed, the number of occurrences of the directions of all spatially similar structures of the retrieval image can be obtained.

[0021] In the above step 7, the direction histogram constructed is H c ={h1,h2,h3,h4}, where h1 is the number of occurrences of the direction θ=0°, h2 is the number of occurrences of the direction θ=45°, h3 is the number of occurrences of the direction θ=90°, and h4 is the number of occurrences of the direction θ=135°.

[0022] Compared with the prior art, the present invention has the following characteristics:

[0023] 1. A feature extraction model that combines the SimCLRV2 self-supervised learning framework with the SwinTransformer network. This not only takes advantage of the feature extraction capabilities of the SwinTransformer network, but also enables the SwinTransformer network to be migrated to tasks with medium or small datasets, making the feature extraction model more adaptable to different tasks and having very strong flexibility and generalization capabilities.

[0024] 2. The SimCLRV2 self-supervised learning framework based on the SwinTransformer encoder is used as the feature extraction model to extract the feature map and depth feature histogram of the image, and construct the direction histogram by finding the spatial similarity structure of the feature map. The depth feature histogram and the direction histogram are used together as the image abstract representation for image retrieval. In this way, the image abstract representation can jointly encode the depth information and spatial information of the image, thereby greatly improving the accuracy of image retrieval. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Schematic diagram of the SimCLRV2 self-supervised learning framework.

[0026] Figure 2 An example of spatially similar structure detection, (a) two-dimensional Gabor filter, (b) feature map and directional feature map, (c) schematic diagram of pixel direction index, (d) local structure with directional sense, and (e) spatially similar structure. DETAILED DESCRIPTION

[0027] 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 with reference to specific examples and the accompanying drawings.

[0028] An image retrieval method that simulates the brain's representation of abstract knowledge includes the following steps:

[0029] Step 1: Use the training image set to train the SimCLRV2 self-supervised learning framework to obtain a trained SimCLRV2 self-supervised learning framework (feature extraction model).

[0030] The training image set of the present invention comes from the famous ImageNet library.

[0031] The principle of the SimCLRV2 self-supervised learning framework is as follows Figure 1As shown, it is built on the basis of the SimCLR learning framework. Since the SimCLRV2 self-supervised learning framework uses only 10% of the labeled data to achieve results that exceed supervised learning, it greatly improves the effect of self-supervised contrastive learning. When selecting the basic encoder of the SimCLRV2 self-supervised learning framework, considering that SwinTransformer provides a more general backbone network for Transformer-based computer vision tasks and can be applied to various computer vision tasks such as image classification, target detection, semantic segmentation, and instance segmentation, the present invention selects the SwinTransformer network of the Swin-L model as the basic encoder of the SimCLRV2 self-supervised learning framework.

[0032] In order to further improve the training effect, in addition to using the original (built-in) expansion image method in the SimCLRV2 self-supervised learning framework to expand the training image set, the present invention also uses a superpixel segmentation method to expand the training image set. When the original expansion image method is used to expand the training image, a transformed image is obtained by random adjustment, scaling or rotation. When the superpixel segmentation method (simple linear iterative clustering, SLIC) is used to expand the training image: first, initialization is performed, and the image is divided into a uniformly distributed initial cluster center grid; then, allocation is performed, and each pixel is assigned to the nearest cluster center based on a distance metric that combines spatial and color information; then, an update is performed, and the cluster center is recalculated based on the average position of all pixels assigned to the cluster; finally, iteration is performed, and the allocation and update are repeated until convergence, that is, the cluster center does not change significantly between iterations; thereby, a transformed image is obtained. The image augmentation method proposed in this paper ensures that the expanded training image set includes both images augmented by the native method, thus maintaining the original advantages of the SimCLRV2 framework, and images augmented by the superpixel method, which increases the difficulty of network learning. By combining these two image augmentation methods, a solid foundation is laid for subsequent image representation and image retrieval, ultimately enabling the trained framework to possess stronger generalization capabilities.

[0033] Step 2: Each retrieval image in the retrieval image database is fed into the trained SimCLRV2 self-supervised learning framework. The trained SimCLRV2 self-supervised learning framework outputs the feature map set Ω and the depth feature histogram H of the retrieval image. d .

[0034] The trained SimCLRV2 self-supervised learning framework has two output ports: one output port is located in the middle of the SimCLRV2 self-supervised learning framework, which is the output port of the basic encoder of the SimCLRV2 self-supervised learning framework, and outputs the feature map set Ω of the retrieval image, which contains 1536 feature maps Ω i , each feature map Ω i The size is 12×12, that is The other output end is located at the end of the trained SimCLRV2 self-supervised learning framework, which is the output end of the entire SimCLRV2 self-supervised learning framework and outputs the deep feature histogram H of the retrieved image. d ={f1,f2,…,f K}. Where K = 1536.

[0035] Step 3: For each feature map Ω in the feature map set of each retrieved image i , use the two-dimensional Gabor filter to perform Gabor filtering on it, such as Figure 2 As shown in (a), the feature map Ω is obtained i Four reference direction feature maps

[0036] Research shows that abstract knowledge is stored in the brain in the form of cognitive maps. If abstract cognitive variables can be represented on a coordinate axis / coordinate system (such as the closeness of kinship, the differences between objects), the excitation positions of the same cell in the brain will not be too far apart, and the preferred positions of the same brain area on the coordinate axis will not be too far apart. In addition, the neural activity space rich in cognitive information enables organisms to make predictions and judgments. In order to facilitate the description of abstract features of images, the present invention uses a two-dimensional Gabor filter to calculate the directional feature map. The Gabor filter is defined as follows:

[0037]

[0038] Where, X = xcosθ + ysinθ, Y = -xsinθ + ycosθ, (x, y) are the coordinates of the pixel point, θ is the reference direction of the filter, γ is the ellipticity, λ is the wavelength, δ is the standard deviation, and φ is the phase. The present invention proposes the reference direction θ = [0°, 45°, 90°, 135°], that is, 4 reference directions, γ = 0.25, λ = 0.56, δ = 2.333, φ = 0 and At this time, each feature map Ω i The reference direction feature maps of 4 reference directions θ = [0°, 45°, 90°, 135°] can be obtained

[0039] Step 4: For each feature map Ω i , all its reference direction feature maps The pixel values ​​of the corresponding pixels are accumulated and assigned as the corresponding pixel values ​​of the directional feature map to obtain the directional feature map s of the feature map i ,like Figure 2 As shown in (b) in the figure, a set of directional feature maps S of each retrieved image is obtained.

[0040] For each feature map Ω i , using its four reference direction feature maps Calculate the directional feature map s i , which is defined as follows:

[0041]

[0042] Among them, ∑{·} represents the accumulation of pixel values ​​of corresponding pixels, that is, for each feature map Ω i , its four reference direction feature maps The pixel values ​​of the pixels at the same coordinates (x, y) are accumulated as the feature map Ω i Directional feature map s i The values ​​of the pixels at coordinates (x, y) are x=1, 2,…, 12 and y=1, 2,…, 12.

[0043] The feature map of each retrieved image is collected After Gabor filtering, according to the above operation, the directional feature map set of each retrieval image can be obtained. K=1536.

[0044] Step 5: For each feature map Ω i , all its reference direction feature maps The direction θ0 corresponding to the pixel maximum value of the corresponding pixel point is used as the assignment of the corresponding pixel point of the direction map to obtain the direction map of the feature map, and then obtain the direction map set O of each retrieval image.

[0045] For each feature map Ω i , using its four reference direction feature maps Calculate the direction map O i , which is defined as follows:

[0046]

[0047] Among them, argmax{·} represents the direction corresponding to the maximum pixel value of the corresponding pixel point, that is, for each feature map Ω i , compared with its four reference direction feature maps The pixel value of the pixel point at the same coordinate (x, y) and the direction θ corresponding to the largest pixel value among the four pixel values ​​are selected as the feature map Ω i Direction map O i The values ​​of the pixels at coordinates (x, y) are x=1, 2,…, 12 and y=1, 2,…, 12.

[0048] The feature map of each retrieved image is collected After Gabor filtering, according to the above operation, the direction map set of each retrieval image can be obtained K=1536.

[0049] Step 6: For each feature map Ω of the feature map set Ω of each retrieved image i , record the direction θ1 of the local structure with directional sense of each pixel; at the same time, for each directional feature map s of the directional feature map set S of each retrieval image i , recording the direction θ2 of the local structure with a sense of direction for each pixel.

[0050] In order to jointly encode depth information and spatial information, the present invention needs to consider the geometric structure characteristics of the feature map set Ω and the directional feature map set S, and consider the dependency between them, that is, to find local structures that appear in the same channel, at the same position, and have the same sense of direction.

[0051] For the feature map Ω i Or directional feature map s i A coordinate in (x o ,y o ) has a pixel value of p0; in a certain direction θ=[0°,45°,90°,135°], there is a pixel with a coordinate of (x o ,y o ) of the pixel point, the two adjacent pixels have pixel values ​​of p o+1 and p o-1 ; If the pixel values ​​of these three pixels are equal, that is, p o-1 =p o =p o+1 , then it is considered that these three pixels form a local structure with a sense of direction, and the direction of the local structure with a sense of direction is θ. Figure 2 (c) in the figure shows a schematic diagram of the pixel direction index. Figure 2 (d) shows a local structure with a sense of direction. i The pixel values ​​of the three adjacent pixels of the current pixel in the 135° direction are the same (shown as the same grayscale value in the figure), so the feature map Ω is determined. iThere is a local structure with a sense of direction, and its direction is 135°; due to the directional feature map s i The pixel values ​​of the three adjacent pixels of the current pixel in the 135° direction are the same (shown as the same grayscale value in the figure), so it is determined that the directional feature map s i There is a local structure with a sense of direction, and its direction is 135°.

[0052] According to the above definition of local structure with directional sense, we can know that: due to the feature map Ω i Or directional feature map s i There are no two adjacent pixels at the edge of the pixel point, so the feature map Ω i Or directional feature map s i It is impossible to form a local structure with a sense of direction at the edge pixel points of Ω. The local structure with a sense of direction will only appear in the feature map Ω. i Or directional feature map s i At non-edge pixels.

[0053] Step 7: Count the number of occurrences of the directions of each spatially similar structure of each retrieval image, and construct a direction histogram using the number of occurrences of the directions of each spatially similar structure recorded in each retrieval image.

[0054] For each retrieval image, the number of feature maps in the feature map set Ω, the number of directional feature maps in the directional feature map set s, and the number of directional maps in the directional map set O are the same (all 1536), and the feature map Ω i , directional feature map s i and direction map O i There is a one-to-one correspondence; each feature map Ω i , directional feature map s i and direction map O i The sizes of are consistent (all 12×12), and the feature map Ω i , feature map s i and direction map O i Pixels at the same coordinates (x, y) are in one-to-one correspondence. On this basis, the number of occurrences of the directions of each spatially similar structure in each search image is calculated as follows:

[0055] If the current feature map Ω of the feature map set Ω i The direction θ1 of the spatial similarity structure of the current pixel point and the corresponding directional feature map s of the directional feature map set S i The direction θ2 of the spatial similarity structure of the corresponding pixel points is the same, and this direction is consistent with the corresponding direction map O of the directional feature map set S. iWhen the values ​​assigned to the corresponding pixel points of θ1 and θ2 are the same, that is, when θ1 = θ2 = θ0, the two directional local structures are considered to be spatially similar structures. In this case, the two directional local structures are considered to be spatially similar structures, and the directions of these two directional local structures are used as the directions of the spatially similar structures. The number of occurrences of the directions of the spatially similar structures is then accumulated by 1. For example, if θ1 = θ2 = θ0 = 45°, the number of occurrences of the direction of 45° is increased by 1. For example, if θ1 = θ2 = θ0 = 90°, the number of occurrences of the direction of 90° is increased by 1. After the traversal is completed, the number of occurrences of the directions of all spatially similar structures in the search image is obtained.

[0056] In the feature map and directional feature map, the spatial similarity structure involves the positioning of channel positions and spatial positions and the association between them. Figure 2 (e) shows the spatial similarity structure, due to the feature map Ω i Or directional feature map s i The corresponding pixels all have a local structure with a sense of direction at 135°. Therefore, it is determined that a spatially similar structure exists at the current pixel, and the direction of the spatially similar structure is 135°. The number of occurrences of the direction 135° is increased by 1. In the present invention, spatially similar structures can present different senses of direction. In addition, it is closely related to the channel information and spatial position information of the depth feature map, and whether they co-occur.

[0057] After the traversal is completed, the direction histogram is constructed using the number of occurrences of each direction recorded in each search image. In the present invention, only the four directions θ = [0°, 45°, 90°, 135°] are considered, so the constructed direction histogram H c ={h1, h2, h3, h4}, where h1 is the number of occurrences of the direction θ=0°, h2 is the number of occurrences of the direction θ=45°, h3 is the number of occurrences of the direction θ=90°, and h4 is the number of occurrences of the direction θ=135°.

[0058] Step 8: The depth feature histogram H of each retrieval image obtained in step 2 is d ={f1, f2, ..., f 1536} and the direction histogram H of each retrieved image obtained in step 7 c = {h1, h2, h3, h4} are concatenated to obtain the multi-tuple feature H of the retrieval image; the multi-tuple feature H of the retrieval image is then normalized to obtain the image abstract representation of the retrieval image

[0059] In constructing abstract representations of images In the process, it includes both the depth features based on the depth information (depth feature histogram H d), and also includes low-level features based on spatial information (direction histogram H c ), through feature integration, the image is abstractly represented It can encode image information more accurately, which is beneficial for subsequent image retrieval.

[0060] Step 9: Perform steps 2 to 8 on the image to be retrieved to obtain an image abstract representation of the image to be retrieved, calculate the L2 distance between the image abstract representation of the image to be retrieved and the image abstract representations of each retrieval image in the retrieval image database, and output the retrieval image with the smallest L2 distance as the retrieval result.

[0061] The image abstract representation descriptor Applied to image retrieval, and uses L2 distance for image matching, that is, only the image abstract representation of the image to be retrieved A search image in the search image database When the L2 distance is less than the set threshold and is the minimum value, the two images are considered to match; otherwise, the two images are considered to not match.

[0062] 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. Therefore, the present invention is not limited to the above-mentioned 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. An image retrieval method that simulates the brain's representation of abstract knowledge, characterized by: The steps are as follows: Step 1: Use the training image set to train the SimCLRV2 self-supervised learning framework to obtain a trained SimCLRV2 self-supervised learning framework; the basic encoder of the SimCLRV2 self-supervised learning framework is the SwinTransformer network; Step 2: Each retrieval image in the retrieval image database is fed into the trained SimCLRV2 self-supervised learning framework, which outputs a feature map set and a deep feature histogram of the retrieval image. Step 3: For each feature map in the feature map set of each search image, perform Gabor filtering on it using a two-dimensional Gabor filter to obtain multiple reference directional feature maps of the feature map; Step 4: For each feature map, the pixel values ​​of the corresponding pixels of all reference directional feature maps are accumulated and used as the values ​​of the corresponding pixels of the directional feature map to obtain the directional feature map of the feature map, and then obtain the directional feature map set of each search image; Step 5: For each feature map, the direction corresponding to the pixel maximum value of the corresponding pixel points of all reference directional feature maps is used as the assignment of the corresponding pixel point of the directional map to obtain the directional map of the feature map, and then obtain the directional map set of each search image; Step 6: For each feature map in the feature map set of each search image, record the direction of the local structure with a sense of direction of each pixel point; at the same time, for each directional feature map in the directional feature map set of each search image, record the direction of the local structure with a sense of direction of each pixel point; Step 7: Count the number of occurrences of the directions of the spatially similar structures of each retrieval image, and construct a direction histogram using the number of occurrences of the directions of the spatially similar structures recorded for each retrieval image: Step 8: Concatenate the depth feature histogram of each retrieval image obtained in step 2 and the direction histogram of each retrieval image obtained in step 7 to obtain a multi-tuple feature of the retrieval image, and then normalize the multi-tuple feature of the retrieval image to obtain an image abstract representation of the retrieval image; Step 9: Perform steps 2 to 8 on the image to be retrieved to obtain an image abstract representation of the image to be retrieved, calculate the L2 distance between the image abstract representation of the image to be retrieved and the image abstract representations of each retrieval image in the retrieval image database, and output the retrieval image with the smallest L2 distance as the retrieval result.

2. The image retrieval method according to claim 1, wherein the method is characterized in that: In step 1, when training the SimCLRV2 self-supervised learning framework, in addition to using the original image expansion method to expand the training image set, the superpixel segmentation method is also used to expand the training image set.

3. The image retrieval method according to claim 1, wherein the method is characterized by: In step 3, after Gabor filtering, each feature map obtains four reference direction feature maps, namely, a 0° reference direction feature map, a 45° reference direction feature map, a 90° reference direction feature map, and a 135° reference direction feature map.

4. The image retrieval method according to claim 1, wherein: In step 6, for a coordinate in the feature map or directional feature map The pixel point, its pixel value is ; in a certain direction The coordinates of The pixel values ​​of the two adjacent pixels are and ; If the pixel values ​​of these three pixels are equal, that is , then it is considered that these three pixels form a local structure with a sense of direction, and the direction of the local structure with a sense of direction is .

5. The image retrieval method according to claim 4, wherein the method is characterized in that: The direction of the local structure with a sense of direction 0°, 45°, 90° or 135°.

6. The image retrieval method according to claim 1, wherein: In step 7, the process of counting the number of occurrences of the direction of the spatially similar structure of each retrieval image is as follows: correspondingly traversing each pixel point of each feature map in the feature map set of the retrieval image and each pixel point of each directional feature map of the directional feature map set of the retrieval image: if the direction of the local structure with a sense of direction of the current pixel point of the current feature map of the feature map set is the same as the direction of the local structure with a sense of direction of the current pixel point of the current directional feature map of the directional feature map set, and the direction of the local structure with a sense of direction is the same as the assignment of the corresponding pixel point of the corresponding directional mapping map of the directional feature map set, then the two local structures with a sense of direction are considered to be spatially similar structures, and the direction of the local structure with a sense of direction is used as the direction of the spatially similar structure, and the number of occurrences of the direction of the spatially similar structure is accumulated by 1; After the traversal is completed, the number of occurrences of directions of all spatially similar structures of the retrieved image can be obtained.

7. The image retrieval method according to claim 1, wherein: In step 7, the constructed direction histogram is ,in For direction The number of occurrences of For direction The number of occurrences of For direction The number of occurrences of For direction The number of occurrences of .

Images