Visual scene recognition method for domain invariant information mining

Abstract

The invention relates to the technical field of robot autonomous navigation and visual identification, and discloses a visual scene identification method for domain invariant information mining, which is realized by forming a DIMNet model by using two parts of a backbone network VGG-DIMM and an aggregation head NetVLAD, and comprises the following steps of: for a given input image, preliminarily extracting a local feature descriptor of the given input image by using the VGG-DIMM; taking the local feature descriptors extracted by the VGG-DIMM as input, and aggregating the local feature descriptors into a one-dimensional global feature descriptor F (I) by using the original NetVLAD; comparing the global feature descriptors F (I) of the query image with the global feature descriptors F (I) of all database images one by one through a similarity algorithm, so as to find out all database images belonging to the same scene as the query image; according to the method, an adopted domain invariant information mining module (DIMM) is included in a cascaded convolution unit group, and the purpose of enhancing the attention of a DIMNet model to a domain invariant static object is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of robot autonomous navigation and visual recognition, and in particular to a visual scene recognition method for mining domain-invariant information. Background Art

[0002] Visual Place Recognition (VPR) technology is an indispensable key technology for the construction of smart cities and national defense security. As a downstream task in the field of image retrieval, it plays an important role in the navigation and positioning systems of robots and unmanned vehicles. The basic process of a VPR system is: given a query image, the most similar image is retrieved from the existing geographic location reference image database, and the retrieved image is used as the estimated geographic location of the query image. However, due to the influence of different perspectives and different environments (such as lighting changes, seasonal changes, etc.), images of the same location show large style differences, which brings great challenges to the VPR task.

[0003] Therefore, how to extract robust descriptors in scene images that can cope with complex environmental changes has become the main starting point for improving the performance of VPR algorithms. However, previous algorithms emphasized the extraction and integration of universal features of images, while ignoring the mining of significant features that play a key role in discriminating VPR tasks. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a visual scene recognition method for domain-invariant information mining. By adopting a domain-invariant information mining module (DIMM), the interdependence between different spatial regions of the feature map is incorporated into the cascaded convolution unit group, so as to achieve the purpose of enhancing the attention of the DIMNet model to domain-invariant static objects.

[0005] The present invention is implemented by the following technical solution: a visual scene recognition method for mining domain-invariant information, which is implemented by using a DIMNet model composed of a backbone network VGG-DIMM and an aggregation head NetVLAD, and includes the following steps:

[0006] 1) For a given input image I, use the backbone network VGG-DIMM to preliminarily extract its local feature descriptors and Where H represents height, W represents width, C represents the number of channels, R represents a real number, and the input image I includes the query image and the database image;

[0007] 2) Local feature descriptors extracted using the backbone network VGG-DIMM As input, NetVLAD uses the original aggregation head Aggregated into a one-dimensional global feature descriptor F (I) , and F (I) ∈R D , where D = cluster_num × C, cluster_num is the number of aggregated clusters in NetVLAD (an optional parameter, preferably set to 64), and C is The number of channels is expressed as follows:

[0008]

[0009] 3) Through the similarity algorithm, the global feature descriptor F of the query image is (I) The global feature descriptor F of all database images (I) A one-to-one comparison is performed to find all database images that belong to the same scene as the query image.

[0010] In order to further better implement the visual scene recognition method of domain-invariant information mining described in the present invention, the following setting is particularly adopted: the similarity algorithm is any one of cosine similarity, Euclidean distance, etc.

[0011] In order to better implement the visual scene recognition method of domain-invariant information mining described in the present invention, the following setting is particularly adopted: the step 1) includes the following specific steps:

[0012] 1.1) Input image I is extracted through VGG16 of the backbone network VGG-DIMM to obtain the basic feature map

[0013] 1.2) After step 1.1), the DIMM of the backbone network VGG-DIMM first increases the capacity of the feature channel information through a 3×3 convolution operation with a step size of 1 to obtain a new feature representation Where r is the feature expansion rate (an optional parameter, preferably set to 4);

[0014] 1.3) To improve The capacity and richness of spatial information, in DIMM, a sliding window of k = 3 × 3 and stride = 1 is used in the new feature representation Traverse along the width and height directions (defined as Expand operation); in particular, when traversing, no convolution operation is performed, only the local receptive field area selected by the sliding window is copied and rearranged according to its original spatial structure to obtain the feature map and The expression is as follows:

[0015]

[0016] In the formula, i and j represent the new feature representation The pixel index in the height and width dimensions, i.e., i = 1, 2, ..., H, j = 1, 2, ..., W; α×k+β is controlled by the size of the sliding window and is the parameter for expanding the local receptive field into columns, representing The index of the last dimension, where α = 0, 1, ..., k-1, β = 1, 2, ..., k;

[0017] 1.4) The attention mask output by the DIMM weight generation module Weight-Block (and The same shape) is fed back to the feature map In order to mine feature maps The important domain invariant information in the model is obtained to obtain a feature expression with strong robustness and generalization ability. Right now:

[0018]

[0019] In the formula, Represents the attention mask output by the weight generation module, i.e., the weight descriptor; The symbols represent the multiplication of corresponding elements;

[0020] 1.5) Steps 1.2) and 1.3) enrich the detail information of the feature map from the two dimensions of channel and space respectively. In order to obtain the relevance of the spatial information of the feature map, the present invention expresses the feature In the last dimension k 2 Perform arithmetic addition operations on the local receptive field to fuse different local receptive field information and obtain

[0021]

[0022] 1.6) In addition, in order to interact with the channel information of the feature map, After further convolution operation with k=3×3 and stride=1 (both k and stride are optional parameters, and both can be set to other parameter values), the local feature descriptor of the final output of the backbone network VGG-DIMM is obtained.

[0023] In order to better implement the visual scene recognition method of domain-invariant information mining described in the present invention, the following setting is particularly adopted: the attention mask The attention map attmap input into the weight generation module is obtained after being processed by a cascaded convolution unit group, where the attention map attmap is a tensor of all 1s with an initialization shape of H×W×4.

[0024] In order to better implement the visual scene recognition method of domain-invariant information mining described in the present invention, the following setting is particularly adopted: the attention mask is defined as:

[0025]

[0026] In order to better realize the visual scene recognition method of domain invariant information mining described in the present invention, the following setting is particularly adopted: the weight generation module includes N convolution units, each of which is composed of a convolution layer Conv, a layer normalization LN and an activation function GELU; wherein the activation function GELU is defined as: Where x represents the feature tensor that needs to be processed by GELU.

[0027] In order to better realize the visual scene recognition method of domain-invariant information mining described in the present invention, the following setting is particularly adopted: the backbone network VGG-DIMM is composed of the Stage1-Stage5 parts of VGG16 and DIMM in series, and the Stage5 part does not contain the last ReLU layer and the layers thereafter in the original Stage5.

[0028] Among them, the VGG part in the backbone network VGG-DIMM can also be replaced by other existing network structures, including but not limited to CNN type, Transformer type or MLP type; the aggregation head NetVLAD can also be replaced by other aggregation heads, including but not limited to GeM Head, MixVPR, etc.

[0029] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0030] (1) The DIMNet model adopted in the present invention can effectively mine domain-invariant information in scene images and enhance the model's discriminability for positive and negative samples. Its performance surpasses the DW-T and NocPlace algorithms newly proposed in 2024, especially on the TokyoTM benchmark, reaching the current best level.

[0031] (2) The DIMNet model proposed in the present invention is a one-stage network with the characteristics of lightweight structure and low complexity, and can be trained end-to-end.

[0032] (3) The DIMM module designed by the present invention is "plug and play" and has good portability. It can be combined with any other model (including but not limited to CNN, Transformer or MLP networks) to function. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is the overall framework diagram of DIMNet.

[0034] Figure 2 This is a structural diagram of a DIMM.

[0035] Figure 3 This is the schematic diagram of the Expand operation. DETAILED DESCRIPTION

[0036] The present invention is further described in detail below in conjunction with examples, but the embodiments of the present invention are not limited thereto.

[0037] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0038] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0039] Glossary:

[0040] DIMNet model: The abbreviation of Domain-invariant Information Mining Network model, that is, domain-invariant information mining network model.

[0041] VGG-DIMM: represents the backbone network consisting of the VGG16 network (Stage1-Stage5 part, and the Stage5 part does not contain the last ReLU layer and subsequent layers in the original Stage5) and DIMM in series.

[0042] DIMM: abbreviation of Domain-invariant Information Mining Module, which means domain-invariant information mining module.

[0043] Conv: convolutional layer.

[0044] Expand: Expand.

[0045] Sum: Add corresponding elements.

[0046] NetVLAD: Aggregation head, a public work.

[0047] Weight-Block: Weight generation module.

[0048] Attention Map: Attention map and attention map attmap refer to the same object.

[0049] First Layer: First layer.

[0050] Middle Layer: Middle layer.

[0051] Last Layer: The last layer.

[0052] Embodiment 1:

[0053] The present invention designs a visual scene recognition method for domain-invariant information mining. The domain-invariant information mining module (DIMM) used incorporates the interdependence between different spatial regions of the feature map in the cascaded convolution unit group, so as to achieve the purpose of enhancing the attention of the DIMNet model to the domain-invariant static objects. The DIMNet model is composed of a backbone network VGG-DIMM and an aggregation head NetVLAD, and includes the following steps:

[0054] 1) For a given input image I, use the backbone network VGG-DIMM to preliminarily extract its local feature descriptors and Where H represents height, W represents width, C represents the number of channels, R represents a real number, and the input image I includes the query image and the database image;

[0055] 2) Local feature descriptors extracted using the backbone network VGG-DIMM As input, NetVLAD uses the original aggregation head Aggregated into a one-dimensional global feature descriptor F (I) , and F (I) ∈R D , where D = cluster_num × C, cluster_num is the number of aggregated clusters in NetVLAD (an optional parameter, preferably set to 64), and C is The number of channels is expressed as follows:

[0056]

[0057] 3) Through the similarity algorithm, the global feature descriptor F of the query image is (I) The global feature descriptor F of all database images (I) A one-to-one comparison is performed to find all database images that belong to the same scene as the query image.

[0058] Embodiment 2:

[0059] This embodiment is further optimized on the basis of the above embodiment, and the similarities with the above technical solution are not repeated here. In order to better realize the visual scene recognition method of domain-invariant information mining described in the present invention, the following setting is particularly adopted: the similarity algorithm is any one of cosine similarity, Euclidean distance, etc.

[0060] Embodiment 3:

[0061] This embodiment is further optimized on the basis of any of the above embodiments, and the same points as the above technical solutions are not repeated here. In order to better realize the visual scene recognition method of domain-invariant information mining described in the present invention, the following setting is particularly adopted: the step 1) includes the following specific steps:

[0062] 1.1) Input image I is extracted through VGG16 of the backbone network VGG-DIMM to obtain the basic feature map

[0063] 1.2) After step 1.1), the DIMM of the backbone network VGG-DIMM first increases the capacity of the feature channel information through a 3×3 convolution operation with a step size of 1 to obtain a new feature representation Where r is the feature expansion rate (an optional parameter, preferably set to 4);

[0064] 1.3) To improve The capacity and richness of spatial information, in DIMM, a sliding window of k = 3 × 3 and stride = 1 is used in the new feature representation Traverse along the width and height directions (defined as Expand operation); in particular, when traversing, no convolution operation is performed, only the local receptive field area selected by the sliding window is copied and rearranged according to its original spatial structure to obtain the feature map and The expression is as follows:

[0065]

[0066] In the formula, i and j represent the new feature representation The pixel index in the height and width dimensions, i.e., i = 1, 2, ..., H, j = 1, 2, ..., W; α×k+β is controlled by the size of the sliding window and is the parameter for expanding the local receptive field into columns, representing The index on the last dimension, where α=0,1,...,k-1, β=1,2,...,k; the principle of the Expand operation is as follows Figure 3 shown.

[0067] 1.4) The attention mask output by the DIMM weight generation module Weight-Block (and The same shape) is fed back to the feature map In order to mine feature maps The important domain invariant information in the model is obtained to obtain a feature expression with strong robustness and generalization ability. Right now:

[0068]

[0069] In the formula, Represents the attention mask output by the weight generation module, i.e., the weight descriptor; The symbols represent the multiplication of corresponding elements;

[0070] 1.5) Steps 1.2) and 1.3) enrich the detail information of the feature map from the two dimensions of channel and space respectively. In order to obtain the relevance of the spatial information of the feature map, the present invention expresses the feature In the last dimension k 2 Perform arithmetic addition operations on the local receptive field to fuse different local receptive field information and obtain

[0071]

[0072] 1.6) In addition, in order to interact with the channel information of the feature map, After further convolution operation with k=3×3 and stride=1 (both k and stride are optional parameters, and both can be set to other parameter values), the local feature descriptor of the final output of the backbone network VGG-DIMM is obtained.

[0073] Embodiment 4:

[0074] This embodiment is further optimized on the basis of any of the above embodiments, and the same points as the above technical solutions are not repeated here. In order to better realize the visual scene recognition method of domain invariant information mining described in the present invention, the following setting is particularly adopted: the attention mask The attention map attmap input into the weight generation module is obtained after being processed by a cascaded convolution unit group, where the attention map attmap is a tensor of all 1s with an initialization shape of H×W×4.

[0075] Embodiment 5:

[0076] This embodiment is further optimized on the basis of any of the above embodiments, and the same points as the above technical solutions are not repeated here. In order to better realize the visual scene recognition method of domain invariant information mining described in the present invention, the following setting is particularly adopted: the attention mask is defined as:

[0077]

[0078] Embodiment 6:

[0079] This embodiment is further optimized on the basis of any of the above embodiments, and the similarities with the above technical solutions are not repeated here. In order to better realize the visual scene recognition method of domain invariant information mining described in the present invention, the following setting is particularly adopted: the weight generation module includes N convolution units, each of which is composed of a convolution layer Conv, a layer normalization LN and an activation function GELU; wherein the activation function GELU is defined as: Where x represents the feature tensor that needs to be processed by GELU.

[0080] Embodiment 7:

[0081] This embodiment is further optimized on the basis of any of the above embodiments, and the similarities with the above technical solutions are not repeated here. In order to better realize the visual scene recognition method of domain-invariant information mining described in the present invention, the following setting is particularly adopted: the backbone network VGG-DIMM is composed of the Stage1-Stage5 part of VGG16 and DIMM in series, and the Stage5 part does not contain the last ReLU layer and the layers thereafter in the original Stage5.

[0082] Among them, the VGG part in the backbone network VGG-DIMM can also be replaced by other existing network structures, including but not limited to CNN type, Transformer type or MLP type; the aggregation head NetVLAD can also be replaced by other aggregation heads, including but not limited to GeM Head, MixVPR, etc.

[0083] Embodiment 8:

[0084] A visual scene recognition method based on domain-invariant information mining, combined with Figure 1 to Figure 3 As shown in the figure, the DIMNet model is implemented by using the backbone network VGG-DIMM (mainly used to perform feature mining in the entire network) and the aggregation head NetVLAD, which includes the following steps:

[0085] Step 1: For a given input image I, use the backbone network VGG-DIMM to preliminarily extract its local feature descriptors and The expression is as follows:

[0086]

[0087] Among them, H represents height, W represents width, C represents the number of channels, R represents a real number, and the input image I includes the query image and the database image.

[0088] The backbone network VGG-DIMM is composed of VGG16 (Stage1-Stage5) and DIMM in series, such as Figure 1 As shown. In particular, in order not to affect the distribution of the processed data, the Stage5 part of VGG16 of the present invention does not include the last ReLU layer and the layers after it of its original structure. The original intention of the design of the present invention is that VGG16 extracts basic features of the input image, and DIMM mines the significant information in the basic features.

[0089] The backbone network VGG-DIMM initially extracts the local feature descriptor of the input image I The following steps are included:

[0090] Step 1.1: Input image I is extracted through VGG16 to obtain the basic feature map

[0091] Step 1.2: After Step 1.1, DIMM (as a plug-and-play feature enhancement module) first increases the capacity of feature channel information through a 3×3 convolution operation with a step size of 1 to obtain a new feature representation Where r is the feature expansion rate (an optional parameter, preferably set to 4).

[0092] Step 1.3: To improve The capacity and richness of spatial information, in DIMM, a sliding window with k = 3 × 3 and stride = 1 is designed in the new feature representation Traverse along the width and height directions (defined as Expand operation); in particular, this process does not perform convolution operations, but only copies the local receptive field area selected by the sliding window and rearranges it according to its original spatial structure to obtain the feature map and The expression is as follows:

[0093]

[0094] In the formula, i and j represent the new feature representation The pixel index in the height and width dimensions, i.e., i = 1, 2, ..., H, j = 1, 2, ..., W; α×k+β is controlled by the size of the sliding window and is the parameter for expanding the local receptive field into columns, representing The index on the last dimension, where α = 0, 1, ..., k-1, β = 1, 2, ..., k; the principle of the Expand operation is as follows Figure 3 shown.

[0095] Step 1.4: Generate the attention mask output by the DIMM weight generation module Weight-Block (and The same shape) is fed back to the feature map In order to mine feature maps The important domain invariant information in the model is obtained to obtain a feature expression with strong robustness and generalization ability. Right now:

[0096]

[0097] In the formula, Represents the attention mask output by the weight generation module, i.e., the weight descriptor; The symbol indicates the multiplication of corresponding elements.

[0098] like Figure 2 As shown in Figure 1, the attention map attmap (initialized as a tensor of all 1s with shape H×W×4) is the input of the weight generation module, which is transformed into an attention mask after being processed by a cascade of convolutional units. It can be acquired through subsequent iterative training The relationship between different spatial regions in the image is analyzed and the salient information that needs to be paid attention to is recalibrated. It can be defined as:

[0099]

[0100] In the weight generation module, each layer can capture the characteristics of the input data. As the number of layers increases, DIMM can learn more complex and advanced data representations. It should be noted that except for the first and last layers, which need to match the channels and sizes of the external feature maps, the structure of the middle layer (N convolution units) of the weight generation module is exactly the same (N is an optional parameter, set to 4 in this example). Specifically, each convolution unit in the weight generation module is composed of a convolution layer Conv (optional, the convolution parameters used in this example are: k = 3×3, stride = 1, dimension L = 4, corresponding to Figure 2 The symbol L in the figure), layer normalization LN and activation function GELU: layer normalization performs better than batch normalization in small batch sample training, and layer normalization is more flexible; the activation function GELU has zero centrality, which helps to alleviate the gradient disappearance problem and can provide better gradient information. It is defined as:

[0101]

[0102] Where x represents the feature tensor that needs to be processed by GELU.

[0103] In general, during the training process, the attention map is iteratively processed by the weight generation module, and can be adaptively obtained during the learning process. The weights of different local features in , and can implicitly capture The semantic relationship between different spatial regions in the image can be used to mine the building information in the image.

[0104] Step 1.5: Step 1.2 and Step 1.3 enrich the detailed information of the feature map from the two dimensions of channel and space respectively. In order to obtain the correlation of the spatial information of the feature map, the present invention expresses the feature In the last dimension k 2 Perform arithmetic addition operations on the local receptive field to fuse different local receptive field information and obtain In addition, in order to exchange the channel information of the feature map, After further convolution operation with k=3×3 and stride=1 (both k and stride are optional parameters, and both can be set to other parameter values), the local feature descriptor of the final output of the backbone network VGG-DIMM is obtained.

[0105] Step 2: Local feature descriptors extracted using the backbone network VGG-DIMM As input, NetVLAD uses the original aggregation head Aggregated into a one-dimensional global feature descriptor F (I) , and F (I) ∈R D , where D = cluster_num × C, cluster_num is the number of aggregated clusters in NetVLAD (an optional parameter, preferably set to 64), and C is The number of channels is expressed as follows:

[0106]

[0107] Step 3: First, apply the operations of Step 1-Step 2 to both the query image and the database image to obtain different global feature descriptors for discrimination. Then, use cosine similarity (Euclidean distance or other similarity calculation methods can also be used) to compare the global feature descriptors of the query image with the global feature descriptors of all database images in the database one by one, so as to find all database images that belong to the same scene as the query image.

[0108] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification or equivalent change made to the above embodiment based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A visual scene recognition method based on domain-invariant information mining, Features: The DIMNet model is implemented by using the backbone network VGG-DIMM and the aggregation head NetVLAD, which includes the following steps: 1) For a given input image I, use the backbone network VGG-DIMM to preliminarily extract its local feature descriptors and Where H represents height, W represents width, C represents the number of channels, R represents a real number, and the input image I includes the query image and the database image; 2) Local feature descriptors extracted using the backbone network VGG-DIMM As input, NetVLAD uses the original aggregation head Aggregated into a one-dimensional global feature descriptor F (I) , and F (I) ∈R D , where D = cluster_num × C, cluster_num is the number of aggregated clusters in NetVLAD, and C is The number of channels is expressed as follows: 3) Through the similarity algorithm, the global feature descriptor F of the query image is (I) The global feature descriptor F of all database images (I) A one-to-one comparison is performed to find all database images that belong to the same scene as the query image.

2. The method for visual scene recognition based on domain-invariant information mining according to claim 1, Features: The similarity algorithm is any one of cosine similarity and Euclidean distance.

3. A visual scene recognition method for mining domain-invariant information according to claim 1 or 2, Features: The step 1) comprises the following specific steps: 1.1) Input image I is extracted through VGG16 of the backbone network VGG-DIMM to obtain the basic feature map 1.2) After step 1.1), the DIMM of the backbone network VGG-DIMM first increases the capacity of the feature channel information through a 3×3 convolution operation with a step size of 1 to obtain a new feature representation Where r is the feature expansion rate; 1.3) In DIMM, a sliding window with k=3×3 and stride=1 is used in the new feature representation Traverse along the width and height directions; when traversing, no convolution operation is performed, only the local receptive field area selected by the sliding window is copied and rearranged according to its original spatial structure to obtain the feature map and The expression is as follows: In the formula, i and j represent the new feature representation The pixel index in the height and width dimensions, i.e., i = 1, 2, ..., H, j = 1, 2, ..., W; α×k+β is controlled by the size of the sliding window and is the parameter for expanding the local receptive field into columns, representing The index of the last dimension, where α = 0, 1, ..., k-1, β = 1, 2, ..., k; 1.4) The attention mask output by the DIMM weight generation module Feedback to feature map In this paper, we obtain a feature expression with strong robustness and strong generalization ability. Right now: In the formula, Represents the attention mask output by the weight generation module, i.e., the weight descriptor; Indicates the multiplication of corresponding elements; 1.5) Express the features In the last dimension k 2 Performing arithmetic addition on 1.6) After further convolution operation, the local feature descriptor of the final output of the backbone network VGG-DIMM is obtained.

4. The method for visual scene recognition based on domain-invariant information mining according to claim 3, Features: The attention mask The attention map attmap input into the weight generation module is obtained after being processed by a cascaded convolution unit group, where the attention map attmap is a tensor of all 1s with an initialization shape of H×W×4.

5. The method for visual scene recognition based on domain-invariant information mining according to claim 3, Features: The attention mask is defined as:

6. The method for visual scene recognition based on domain-invariant information mining according to claim 3, Features: The weight generation module includes N convolution units, each of which is composed of a convolution layer Conv, a layer normalization LN and an activation function GELU; wherein the activation function GELU is defined as: Where x represents the feature tensor that needs to be processed by GELU.

7. A visual scene recognition method for mining domain-invariant information according to claim 1, 2, 4, 5 or 6, Features: The backbone network VGG-DIMM is composed of the Stage1-Stage5 parts of VGG16 and DIMM in series, and the Stage5 part does not contain the last ReLU layer and the layers thereafter in the original Stage5.

8. The method for visual scene recognition based on domain-invariant information mining according to claim 3, Features: The backbone network VGG-DIMM is composed of the Stage1-Stage5 parts of VGG16 and DIMM in series, and the Stage5 part does not contain the last ReLU layer and the layers thereafter in the original Stage5.

Images