Fast image matching method based on deep learning

Abstract

The present invention relates to the field of image processing technology, and in particular to a method for fast image matching based on deep learning, comprising: establishing an image database for image matching, and preprocessing different images therein; extracting image salient features at each scale for each image in the image database; establishing an image feature retrieval tree for each image in the image database using local salient features of each image in the image database at the original scale and at each size; and fast matching of images using a deep learning algorithm combined with an image matching feature retrieval forest, and providing feedback on the matching results of the images. The present invention reduces the amount of calculation in the image matching process and makes the overall matching speed faster.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a fast image matching method based on deep learning. Background Art

[0002] With the development of science and technology, image processing technology is more and more widely used in our lives, especially image matching technology is a hot topic. Image matching technology is widely used in fields such as object recognition and target detection, image retrieval and similar image search, as well as industrial automation and quality control.

[0003] Traditional methods (such as image-based similarity matching) usually involve multiple stages of feature extraction, feature description, and matching during image matching, which are computationally complex and time-consuming. In addition, for complex image scenes, traditional methods may not be able to provide accurate and efficient matching results. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a fast image matching method based on deep learning, the method comprising:

[0005] Obtaining an image to be matched and an image database;

[0006] Obtain the local salient features of each image in the image database at the original scale and at each size;

[0007] According to the local significant features of each image in the image database at the original scale, all image feature retrieval trees of the image database and the first node of each image feature retrieval tree are obtained; according to the first node of each image feature retrieval tree, the first branch of the second node of each image feature retrieval tree is obtained; according to the first branch of the second node of each image feature retrieval tree, the third node of each image feature retrieval tree is obtained; according to the third node of each image feature retrieval tree, all complete image feature retrieval trees in the image database are obtained; according to all complete image feature retrieval trees in the image database, an image matching feature retrieval forest of the image database is constructed;

[0008] A deep learning algorithm is used in combination with the image matching feature retrieval forest of the image database to complete the fast matching of the image to be matched, and the image matching confidence of the image to be matched is obtained; the image matching result is fed back according to the image matching confidence of the image to be matched.

[0009] Preferably, the method of obtaining the local significant features of each image in the image database at the original scale and at each size includes:

[0010] For the i-th image in the image database, first, the SIFT method is used to obtain the local descriptor corresponding to the i-th image, and the local descriptor is used as the local salient feature of the i-th image at the original scale; secondly, the scale transformation algorithm is used to transform the scale of the i-th image to obtain the image at each scale of the i-th image; finally, the SIFT method is used to obtain the local descriptor corresponding to the image at each scale of the i-th image, and the local descriptor is used as the local salient feature of the i-th image at each scale.

[0011] Preferably, the method of obtaining all image feature retrieval trees of the image database and the first node of each image feature retrieval tree according to the local significant features of each image in the image database at the original scale includes the following specific methods:

[0012] The local significant features of each image in the image database at the original scale are clustered using the adaptive K-means algorithm to obtain M clusters; each cluster contains several images in the image database, each cluster corresponds to an image feature retrieval tree, and the mth cluster is used as the first node of the mth image feature retrieval tree.

[0013] Preferably, the method of obtaining the first branch of the second node of each image feature retrieval tree according to the first node of each image feature retrieval tree includes:

[0014] For the mth image feature retrieval tree corresponding to the mth cluster among the M clusters, the adaptive LOF algorithm is used to cluster the first node of the mth image feature retrieval tree to obtain clusters; finally, the local significant features of several images corresponding to the cluster at the second scale are used as the first branch of the second node of the mth image feature retrieval tree.

[0015] Preferably, the method of obtaining the third node of each image feature retrieval tree according to the first branch of the second node of each image feature retrieval tree includes:

[0016] For the mth image feature retrieval tree corresponding to the mth cluster among the M clusters, the adaptive LOF algorithm is used to cluster the local significant features of all images in the image database at the third scale corresponding to the first branch of the second node of the mth image feature retrieval tree to obtain clusters and multiple outliers; finally, the local significant features of several images corresponding to the cluster at the third scale are used as the first branch of the third node of the mth image feature retrieval tree; the local significant features of several images corresponding to the multiple outliers at the third scale are used as the second branch of the third node of the mth image feature retrieval tree, and the first branch of the third node and the second branch of the third node constitute the third node.

[0017] Preferably, the specific method of obtaining all complete image feature retrieval trees in the image database according to the third node of each image feature retrieval tree includes:

[0018] The first branch of the third node of the m-th image feature retrieval tree is clustered with the local significant features of all images in the image database at the fourth scale by using the adaptive LOF algorithm to obtain clusters and multiple outliers; finally, the local significant features of several images at the fourth scale corresponding to the clusters are used as the first branch of the third node of the m-th image feature retrieval tree; the local significant features of several images at the fourth scale corresponding to the multiple outliers are used as the second branch of the third node of the m-th image feature retrieval tree, and the first branch of the third node and the second branch of the third node constitute the third node;

[0019] According to the method of constructing the third node of the mth image feature retrieval tree, all remaining nodes of the mth image feature retrieval tree are obtained until the number of images in the obtained cluster is equal to 1, and then the complete mth image feature retrieval tree is obtained; wherein, the setting process of all remaining nodes of the mth image feature retrieval tree is the same as the construction of the third node of the mth image feature retrieval tree.

[0020] Preferably, the method of using a deep learning algorithm in combination with an image database to retrieve an image matching feature forest to complete the rapid matching of the images to be matched includes the following specific methods:

[0021] First, the local salient features of the image to be matched at the original scale are obtained by the SIFT method; secondly, the local salient features of the image to be matched at the original scale are classified with the first nodes of all image matching feature retrieval trees in the image matching feature retrieval forest by deep learning, and the image matching feature retrieval tree corresponding to the image to be matched is obtained and recorded as the first image matching feature retrieval tree; then, the local salient features of the image to be matched at the second scale are obtained by the SIFT method; the local salient features of the image to be matched at the original scale are classified with the first branch of the second node of the first image matching feature retrieval tree by deep learning, and if the results belong to the same category, the first branch of the third node of the first image matching feature retrieval tree is classified;

[0022] And so on, until the local salient features of the image to be matched at the nth scale and all the local salient features of the first branch in the nth node in the first image matching feature retrieval tree cannot be classified into one category using the deep learning algorithm, the image to be matched is classified into the first branch in the n-1th node in the first image matching feature retrieval tree; then the fast matching of the image to be matched is completed.

[0023] Preferably, the obtaining of the image matching confidence of the image to be matched includes the following specific methods:

[0024] Obtain the node credibility from the first node to the nth node of the first image matching feature retrieval tree. The calculation expression of the node credibility is:

[0025]

[0026] Among them, G n' represents the node credibility of the n'th node of the first image matching feature retrieval tree; N represents the total number of nodes in the first image matching feature retrieval tree;

[0027] The accumulated result of the node credibility of the first image matching feature retrieval tree from the first node to the nth node is used as the node credibility of the image matching feature retrieval tree of the image to be matched; the product of the image similarity between the image to be matched and the matching image of the image to be matched and the node credibility of the image matching feature retrieval tree of the image to be matched is used as the image matching confidence of the image to be matched.

[0028] Preferably, the method for acquiring the matching image of the image to be matched is:

[0029] When the local salient features obtained at the nth scale of the image to be matched and all the local salient features in the nth node in the first image matching feature retrieval tree cannot be classified into one category using the deep learning algorithm, the image similarity is calculated for each image in the image database corresponding to all the local salient features in the n-1th node in the first image matching feature retrieval tree, and then the image in the n-1th node with the greatest image similarity to the image to be matched is selected as the matching image of the image to be matched.

[0030] Preferably, the method of providing feedback on the image matching result according to the image matching confidence of the image to be matched includes the following specific methods:

[0031] When the image matching confidence of the image to be matched is less than the preset threshold β, there is no corresponding matching image between the image to be matched and all the images in the image database; when the image matching confidence of the image to be matched is greater than or equal to the preset threshold β, the matching image of the image to be matched is used as the image matching output result of the image to be matched.

[0032] The beneficial effects of the technical solution of the present invention are as follows: the traditional image matching method usually involves multiple stages of feature extraction, feature description and matching, which are complex in calculation and time-consuming. For complex image scenes, the traditional method may not be able to provide accurate and efficient matching results. The present invention extracts local significant features at different scales for different images in the image database, and then uses different scales to establish corresponding image matching feature retrieval trees, which are used to match images. In the overall matching process, the quantification of local significant features of images at different scales is used, so that the accuracy of the matching process is higher. In the matching process, the overall node trend of the established image matching feature retrieval tree is that the number of images is decreasing, so that the amount of calculation in the image matching process is decreasing, and the overall matching speed is faster. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 This is a flowchart of the steps of the deep learning-based image fast matching method of the present invention. DETAILED DESCRIPTION

[0035] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the specific implementation method, structure, features and effects of the image fast matching method based on deep learning proposed by the present invention in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0036] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0037] The specific scheme of the deep learning-based image fast matching method provided by the present invention is described in detail below with reference to the accompanying drawings.

[0038] See also Figure 1 , which shows a flowchart of a method for fast image matching based on deep learning provided by an embodiment of the present invention, the method comprising the following steps:

[0039] Step S001: Establish an image database for image matching and pre-process different images therein.

[0040] It should be noted that the scenario targeted by this embodiment is to quickly match the image P to be matched with the image in the database, so it is necessary to establish a corresponding image database; the specific scenario can obtain the image of the corresponding scene according to the actual needs of the implementer, for example, when image matching is used for face recognition, the corresponding image collected is a face image, and when image matching is used for workpiece defect detection, the corresponding image collected is a workpiece image; because each collected image has different sources and acquisition devices, there is a little difference in conventional features between different images, so it is necessary to preprocess it so that all images are of the same size, and when extracting features of images at different scales, the acquisition scales of different images can be synchronized to avoid different nodes of different retrieval trees due to different scales; the multi-channel image is grayed to make it a single channel, so as to reduce the amount of calculation of subsequent calculation formulas.

[0041] Specifically, in order to implement the deep learning-based fast image matching method proposed in this embodiment, it is necessary to first establish an image database. The specific process is: collect the images to be matched, collect all images in any specific scene, and use them as the basic data set for subsequent image matching; the preprocessing method in this embodiment is mainly: use downsampling and upsampling algorithms to crop and grayscale the images to obtain all preprocessed images; use a database management system to establish an image database for all preprocessed images.

[0042] So far, the image database is obtained through the above method.

[0043] Step S002: extracting image salient features at each scale for each image in the image database.

[0044] It should be noted that when using the image to be matched with different images in the image database, the existing algorithm uses the features of the image to be matched to analyze and decompose the features of all images in the image database, and then matches the image by matching the features of the image to be matched with all images in the image database. The overall image matching accuracy is not high, and because a large amount of calculation is required, the matching time is relatively long. In this embodiment, in order to ensure the accuracy of image matching, the image salient features at different scales are extracted for different images in the image database.

[0045] Specifically, taking the i-th image in the image database as an example, first, the SIFT method is used to obtain the local descriptor corresponding to the i-th image, and the local descriptor is used as the local salient feature of the i-th image at the original scale; secondly, the scale transformation algorithm is used to transform the scale of the i-th image to obtain the image at each scale of the i-th image; finally, the SIFT method is used to obtain the local descriptor corresponding to the image at each scale of the i-th image, and the local descriptor is used as the local salient feature of the i-th image at each scale; wherein the SIFT method and the scale transformation algorithm are prior arts and will not be described in detail here.

[0046] Similarly, all images in the image database are processed to obtain local salient features of each image in the image database at the original scale and at each size.

[0047] So far, the local salient features of each image in the image database at the original scale and at each size are obtained through the above method.

[0048] Step S003: Establish an image feature retrieval tree for each image in the image database using the local significant features of each image in the image database at the original scale and at each size.

[0049] It should be noted that in order to reduce the amount of calculation during image matching and improve the matching accuracy, and to make different images with similar image salient features at the same scale in the same image matching feature retrieval tree as much as possible, different image matching feature retrieval trees are established by utilizing the image salient features at different scales, so as to achieve fast image matching.

[0050] 1. Obtain the image feature retrieval tree corresponding to each image in the image database, and the first node corresponding to each image feature retrieval tree.

[0051] It should be noted that, for the image to be matched and all the I images in the overall database, the present invention performs precision matching based on macro to detail, so when each tree is established, its corresponding first node must be at the macro level.

[0052] The local significant features of the image obtained are obtained by capturing the most significant local features of the image at the current scale. When matching images, in order to reduce the overall amount of calculation, the feature points of the image are generally selected for corresponding matching. Therefore, the present embodiment can use the adaptive K-means algorithm to match the local significant features of the image at the original scale corresponding to all images, and obtain several clusters, wherein the local significant features of the image in each cluster are obviously similar, that is, the corresponding images have certain similarities. When matching any image with an image in a database, it is only necessary to continue to obtain the local significant features of the image corresponding to the original scale of the image to be matched, and use the deep learning algorithm to classify the image significance corresponding to the first node of different trees, so as to exclude most of the images that are not similar to the image to be matched at the macro level, wherein the majority refers to images that are not on the same tree, and only the image to be matched is matched with the image on the corresponding tree; thereby reducing the overall amount of matching calculation and improving the speed of image matching.

[0053] Specifically, the local significant features of each image in the image database at the original scale are clustered using an adaptive K-means algorithm to obtain M clusters; each cluster corresponds to an image feature retrieval tree, each cluster contains several images in the image database, and the total number of image feature retrieval trees in the image database is M.

[0054] Taking the mth cluster among M clusters as an example, the mth cluster is used as the first node of the mth image feature retrieval tree.

[0055] At this point, all the image feature retrieval trees of the image database and the first node of each image feature retrieval tree are obtained.

[0056] 2. Build a complete image feature retrieval tree.

[0057] It should be noted that the local salient features corresponding to the first node of each image feature retrieval tree are the salient features extracted from the image in the image database without scale change, which reflect the more significant local features of the corresponding image as a whole, that is, the salient features at the macro level. When performing image matching, the image to be matched and some images in the image database only have a certain similarity at the macro level, but there are large differences in details. Therefore, this embodiment uses the salient features of images at different scales in each image feature retrieval tree to construct the remaining child nodes, so as to avoid the above situation during the image matching process.

[0058] In order to perform detailed analysis on the second scale through all images corresponding to each image feature retrieval tree, the accuracy of image matching can be improved by using details. The essential purpose of branching each image feature retrieval tree is to ensure the amount of matching data under the condition of ensuring a certain image matching accuracy. The purpose of using clustering algorithm to make outliers and feature similarities as different branches is that if the image to be matched is matched in the mth image feature retrieval tree, its corresponding local significant features at the second scale belong to the same category as the first branch in the mth image feature retrieval tree, then it means that the image to be matched has a high similarity with most of the images in the mth image feature retrieval tree, and it can also be matched with higher accuracy. If it belongs to the second branch, it means that it has a high similarity with a small number of images in the mth image feature retrieval tree, and its corresponding details are relatively more unique. Therefore, when performing a higher-precision match, it is necessary to analyze different images separately, which requires a larger amount of calculation, thereby reducing the overall matching speed. Among them, the branching of nodes is essentially a pruning operation to reduce the amount of calculation in the matching process. As the node increases, the total number of images in the node will decrease after each branch, and the overall matching accuracy will be improved within a certain range. The less the amount of calculation, the faster the matching speed will be.

[0059] Specifically, for the mth image feature retrieval tree corresponding to the mth cluster among the M clusters, the specific method of constructing the second node of the mth image feature retrieval tree is:

[0060] Firstly, the adaptive LOF algorithm is used to cluster the local significant features of all images in the image database at the second scale corresponding to the m-th image feature retrieval tree to obtain clusters and multiple outliers; finally, the local significant features of several images corresponding to the clusters at the second scale are used as the first branch of the second node of the m-th image feature retrieval tree; the local significant features of several images corresponding to the multiple outliers at the second scale are used as the second branch of the second node of the m-th image feature retrieval tree, and the first branch of the second node and the second branch of the second node constitute the second node.

[0061] The specific method of constructing the third node of the mth image feature retrieval tree is:

[0062] Firstly, the adaptive LOF algorithm is used to cluster the local significant features of all images in the image database at the third scale corresponding to the first branch of the second node of the m-th image feature retrieval tree to obtain clusters and multiple outliers; finally, the local significant features of several images at the third scale corresponding to the clusters are used as the first branch of the third node of the m-th image feature retrieval tree; the local significant features of several images at the third scale corresponding to multiple outliers are used as the second branch of the third node of the m-th image feature retrieval tree, and the first branch of the third node and the second branch of the third node constitute the third node.

[0063] By constructing the third node of the mth image feature retrieval tree, all remaining nodes of the mth image feature retrieval tree are obtained until the number of images in the obtained cluster is equal to 1, and then the clusters cannot be clustered using the adaptive LOF algorithm; and then the complete mth image feature retrieval tree is obtained. The setting process of all remaining nodes of the mth image feature retrieval tree is the same as the construction of the third node of the mth image feature retrieval tree.

[0064] At this point, each complete image feature retrieval tree in the image database is obtained.

[0065] All complete image feature retrieval trees in the image database constitute an image matching feature retrieval forest of the image database.

[0066] At this point, the image matching feature retrieval forest of the image database is obtained through the above method.

[0067] Step S004: Use a deep learning algorithm combined with an image matching feature retrieval forest to quickly match images, and provide feedback on the image matching results.

[0068] 1. Use deep learning algorithm combined with image matching feature retrieval forest to quickly match the images to be matched.

[0069] Specifically, the specific process of using deep learning algorithm combined with image matching feature retrieval forest to perform fast image matching is as follows:

[0070] First, the SIFT method is used to obtain the local salient features of the image to be matched at the original scale; secondly, deep learning is used to classify the local salient features of the image to be matched at the original scale and the first nodes of all image matching feature retrieval trees in the image matching feature retrieval forest, and the image matching feature retrieval tree corresponding to the image to be matched is obtained and recorded as the first image matching feature retrieval tree; then, the SIFT method is used to obtain the local salient features of the image to be matched at the second scale; deep learning is used to classify the local salient features of the image to be matched at the original scale and the first branch of the second node of the first image matching feature retrieval tree. If the results belong to the same category, the first branch of the third node of the first image matching feature retrieval tree is classified; if the results do not belong to the same category, no subsequent processing is performed and the matching is stopped.

[0071] In this way, the local salient features of the image to be matched at different scales are classified with the corresponding nodes of the first image matching feature retrieval tree using a deep learning algorithm, until the local salient features of the image to be matched at the nth scale and all the local salient features of the first branch in the nth node in the first image matching feature retrieval tree cannot be classified into one category using the deep learning algorithm, and the image to be matched is classified into the first branch in the n-1th node in the first image matching feature retrieval tree; and the image to be matched and all the local salient features in the n-1th node in the first image matching feature retrieval tree correspond to each image in the image database, and the image similarity is calculated using the Structural Similarity Index (SSIM), and then the image in the n-1th node with the greatest image similarity to the image to be matched is selected as the matching image of the image to be matched.

[0072] At this point, the fast matching of the image to be matched is completed.

[0073] 2. Obtain the image matching confidence of the image to be matched.

[0074] Among them, the node credibility of the first image matching feature retrieval tree from the first node to the nth node is obtained, and the calculation expression of the node credibility is:

[0075]

[0076] Among them, G n' represents the node credibility of the n'th node of the first image matching feature retrieval tree; N represents the total number of nodes in the first image matching feature retrieval tree.

[0077] The accumulated result of the node credibility of the first image matching feature retrieval tree from the first node to the nth node is used as the node credibility of the image matching feature retrieval tree of the image to be matched; the product of the image similarity between the image to be matched and the matching image of the image to be matched and the node credibility of the image matching feature retrieval tree of the image to be matched is used as the image matching confidence of the image to be matched.

[0078] At this point, the image matching confidence of the image to be matched is obtained.

[0079] 3. Provide feedback on the image matching results based on the image matching confidence of the image to be matched.

[0080] A threshold value β is preset, wherein this embodiment is described by taking β=0.6 as an example, and this embodiment is not specifically limited, wherein β is determined according to specific implementation conditions.

[0081] When the image matching confidence of the image to be matched is less than the preset threshold β, it means that there is no corresponding matching image between the image to be matched and all the images in the image database; when the image matching confidence of the image to be matched is greater than or equal to the preset threshold β, the matching image of the image to be matched is used as the image matching output result of the image to be matched.

[0082] At this point, this embodiment is completed.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A fast image matching method based on deep learning, characterized in that: The method comprises the following steps: Obtaining an image to be matched and an image database; Obtain the local salient features of each image in the image database at the original scale and at each size; According to the local significant features of each image in the image database at the original scale, all image feature retrieval trees of the image database and the first node of each image feature retrieval tree are obtained; according to the first node of each image feature retrieval tree, the first branch of the second node of each image feature retrieval tree is obtained; according to the first branch of the second node of each image feature retrieval tree, the third node of each image feature retrieval tree is obtained; according to the third node of each image feature retrieval tree, all complete image feature retrieval trees in the image database are obtained; according to all complete image feature retrieval trees in the image database, an image matching feature retrieval forest of the image database is constructed; The deep learning algorithm is combined with the image matching feature retrieval forest of the image database to complete the fast matching of the image to be matched, and the image matching confidence of the image to be matched is obtained; the matching result of the image is fed back according to the image matching confidence of the image to be matched; The specific method of obtaining all image feature retrieval trees of the image database and the first node of each image feature retrieval tree according to the local significant features of each image in the image database at the original scale includes: The local significant features of each image in the image database at the original scale are clustered using the adaptive K-means algorithm to obtain M clusters; each cluster contains several images in the image database, each cluster corresponds to an image feature retrieval tree, and the mth cluster is used as the first node of the mth image feature retrieval tree; The specific method of obtaining the first branch of the second node of each image feature retrieval tree according to the first node of each image feature retrieval tree includes: For the mth image feature retrieval tree corresponding to the mth cluster among the M clusters, the first node of the mth image feature retrieval tree is clustered using the adaptive LOF algorithm to obtain a cluster; finally, the local significant features of several images corresponding to the cluster at the second scale are used as the first branch of the second node of the mth image feature retrieval tree; The specific method of obtaining the third node of each image feature retrieval tree according to the first branch of the second node of each image feature retrieval tree is as follows: For the m-th image feature retrieval tree corresponding to the m-th cluster among the M clusters, the first branch of the second node of the m-th image feature retrieval tree corresponding to the local significant features of all images in the image database at the third scale is clustered using the adaptive LOF algorithm to obtain clusters and multiple outliers; finally, the local significant features of several images corresponding to the cluster at the third scale are used as the first branch of the third node of the m-th image feature retrieval tree; the local significant features of several images corresponding to the multiple outliers at the third scale are used as the second branch of the third node of the m-th image feature retrieval tree, and the first branch of the third node and the second branch of the third node constitute the third node; The specific method of obtaining all complete image feature retrieval trees in the image database according to the third node of each image feature retrieval tree is as follows: The first branch of the third node of the m-th image feature retrieval tree is clustered with the local significant features of all images in the image database at the fourth scale by using the adaptive LOF algorithm to obtain clusters and multiple outliers; finally, the local significant features of several images at the fourth scale corresponding to the clusters are used as the first branch of the third node of the m-th image feature retrieval tree; the local significant features of several images at the fourth scale corresponding to the multiple outliers are used as the second branch of the third node of the m-th image feature retrieval tree, and the first branch of the third node and the second branch of the third node constitute the third node; According to the method of constructing the third node of the mth image feature retrieval tree, all remaining nodes of the mth image feature retrieval tree are obtained until the number of images in the obtained cluster is equal to 1, and then the complete mth image feature retrieval tree is obtained; wherein, the setting process of all remaining nodes of the mth image feature retrieval tree is the same as the construction of the third node of the mth image feature retrieval tree.

2. The image fast matching method based on deep learning according to claim 1, characterized in that: The specific method of obtaining the local significant features of each image in the image database at the original scale and at each size includes: For the i-th image in the image database, first, the SIFT method is used to obtain the local descriptor corresponding to the i-th image, and the local descriptor is used as the local salient feature of the i-th image at the original scale; secondly, the scale transformation algorithm is used to transform the scale of the i-th image to obtain the image at each scale of the i-th image; finally, the SIFT method is used to obtain the local descriptor corresponding to the image at each scale of the i-th image, and the local descriptor is used as the local salient feature of the i-th image at each scale.

3. The image fast matching method based on deep learning according to claim 1, characterized in that: The method of using a deep learning algorithm combined with an image database to retrieve an image matching feature forest to complete the rapid matching of the image to be matched includes the following specific methods: First, the local salient features of the image to be matched at the original scale are obtained by the SIFT method; secondly, the local salient features of the image to be matched at the original scale are classified with the first nodes of all image matching feature retrieval trees in the image matching feature retrieval forest by deep learning, and the image matching feature retrieval tree corresponding to the image to be matched is obtained and recorded as the first image matching feature retrieval tree; then, the local salient features of the image to be matched at the second scale are obtained by the SIFT method; the local salient features of the image to be matched at the original scale are classified with the first branch of the second node of the first image matching feature retrieval tree by deep learning, and if the results belong to the same category, the first branch of the third node of the first image matching feature retrieval tree is classified; And so on, until the local salient features of the image to be matched at the nth scale and all the local salient features of the first branch in the nth node in the first image matching feature retrieval tree cannot be classified into one category using the deep learning algorithm, the image to be matched is classified into the first branch in the n-1th node in the first image matching feature retrieval tree; then the fast matching of the image to be matched is completed.

4. The method for fast image matching based on deep learning according to claim 3, characterized in that: The specific method of obtaining the image matching confidence of the image to be matched includes: Obtain the node credibility from the first node to the nth node of the first image matching feature retrieval tree. The calculation expression of the node credibility is: Among them, G n' represents the node credibility of the n'th node of the first image matching feature retrieval tree; N represents the total number of nodes in the first image matching feature retrieval tree; The accumulated result of the node credibility of the first image matching feature retrieval tree from the first node to the nth node is used as the node credibility of the image matching feature retrieval tree of the image to be matched; the product of the image similarity between the image to be matched and the matching image of the image to be matched and the node credibility of the image matching feature retrieval tree of the image to be matched is used as the image matching confidence of the image to be matched.

5. The method for fast image matching based on deep learning according to claim 4, characterized in that: The method for obtaining the matching image of the image to be matched is: When the local salient features obtained at the nth scale of the image to be matched and all the local salient features in the nth node in the first image matching feature retrieval tree cannot be classified into one category using the deep learning algorithm, the image similarity is calculated for each image in the image database corresponding to all the local salient features in the n-1th node in the first image matching feature retrieval tree, and then the image in the n-1th node with the greatest image similarity to the image to be matched is selected as the matching image of the image to be matched.

6. The method for fast image matching based on deep learning according to claim 1, characterized in that: The method of providing feedback on the image matching result according to the image matching confidence of the image to be matched includes: When the image matching confidence of the image to be matched is less than the preset threshold β, there is no corresponding matching image between the image to be matched and all the images in the image database; when the image matching confidence of the image to be matched is greater than or equal to the preset threshold β, the matching image of the image to be matched is used as the image matching output result of the image to be matched.

Images