Cross-age face recognition method

Abstract

The invention relates to the technical field of face recognition, and particularly discloses a cross-age face recognition method, which is characterized in that two cross-age face images are processed by using an image processing technology based on deep learning, face structure feature representations of the two images are mined, the ages of the two images are predicted based on the face structure feature representations, and meanwhile, the face structure feature representations of the two images are extracted. And performing difference comparison analysis on the face structure features of the cross-age face image, and further predicting an age gap based on the face difference features, so as to determine whether the face objects in the cross-age face image are the same object based on a dual-channel prediction result. Through the two-channel contrast learning mechanism, the problem that the recognition accuracy of a traditional face recognition technology in a cross-age scene is reduced can be effectively solved, and the accuracy and robustness of face recognition are improved.

Description

Technical Field

[0001] The present application relates to the field of face recognition technology, and more specifically, to a cross-age face recognition method. Background Art

[0002] Face recognition technology has shown high application value in many fields, such as access control security, security permissions, etc. However, as time goes by, the appearance of the face will change, especially the facial details and texture, and there are problems such as skin looseness and skin color changes.

[0003] In existing face recognition technology, training face classification is the main method, hoping to learn the similarities and differences in face distribution during the training process, so that the model has the ability to cluster the same faces and distinguish different faces. When facing cross-age face recognition, due to the large difference in face data under the label, it is difficult for the model to optimize in the desired direction during the learning process, resulting in a decrease in face recognition accuracy when applied.

[0004] Therefore, a cross-age face recognition method is expected. Summary of the invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a cross-age face recognition method, which uses deep learning-based image processing technology to process two cross-age face images, mines the facial structure feature representations of the two, and predicts the ages of the two based on this. At the same time, the facial structure features of the cross-age face images are subjected to difference comparison analysis, and the age difference is further predicted based on the facial difference features, so as to determine whether the face objects in the cross-age face images are the same object based on the prediction results of the dual channels. Through this dual-channel comparative learning mechanism, the problem of decreased recognition accuracy of traditional face recognition technology in cross-age scenarios can be effectively solved, and the accuracy and robustness of face recognition can be improved.

[0006] Accordingly, according to one aspect of the present application, a cross-age face recognition method is provided, which includes:

[0007] Acquire a first face image to be recognized and a second face image to be recognized;

[0008] Performing age prediction and difference calculation on the first face image to be recognized and the second face image to be recognized to obtain an age difference anchor value;

[0009] Performing facial feature difference measurement on the first to-be-recognized face image and the second to-be-recognized face image to obtain a first-to-second to-be-recognized face semantic difference feature map;

[0010] Performing attention enhancement on the first and second face semantic difference feature maps of the objects to be identified to obtain the first and second face semantic difference enhanced feature maps of the objects to be identified;

[0011] Based on the semantic difference enhanced feature map of the first and second objects to be identified, generating an age difference decoding value based on facial semantics;

[0012] In response to the difference between the age difference decoding value based on facial semantics and the age difference anchor value being within a preset range, it is determined that the objects in the first face image to be recognized and the second face image to be recognized are the same object.

[0013] In the above-mentioned cross-age face recognition method, attention is enhanced on the first and second semantic differential feature maps of the faces of the objects to be identified, including: performing adaptive attenuation modulation based on spatial distribution guidance on the pixel-level semantic association structure of the first and second semantic differential feature maps of the faces of the objects to be identified to obtain a semantic association weight map between spatial attenuation pixels of the semantic differential features of the faces of the objects to be identified; based on the semantic association weight map between spatial attenuation pixels of the semantic differential features of the faces of the objects to be identified, feature modulation is performed on the first and second semantic differential feature maps of the faces of the objects to be identified to obtain the first and second semantic differential enhanced feature maps of the faces of the objects to be identified.

[0014] In the above-mentioned cross-age face recognition method, age prediction and difference calculation are performed on the first face image to be recognized and the second face image to be recognized to obtain an age difference anchor value, including: inputting the first face image to be recognized and the second face image to be recognized into a face feature extractor based on a void convolutional neural network model respectively to obtain a first face feature map of the object to be recognized and a second face feature map of the object to be recognized; inputting the first face feature map of the object to be recognized and the second face feature map of the object to be recognized into an age regression predictor based on a decoder respectively to obtain a first age prediction value and a second age prediction value; calculating the difference between the first age prediction value and the second age prediction value to obtain the age difference anchor value.

[0015] In the above-mentioned cross-age face recognition method, facial feature differences are measured for the first face image to be identified and the second face image to be identified to obtain a first-second face semantic differential feature map of the objects to be identified, including: inputting the first face feature map of the object to be identified and the second face feature map of the object to be identified into a face semantic difference feature measurement network to obtain the first-second face semantic differential feature map of the objects to be identified, wherein the face semantic difference feature measurement network is used to calculate the positional difference between the first face feature map of the object to be identified and the second face feature map of the object to be identified to obtain the first-second face semantic differential feature map of the objects to be identified.

[0016] In the above-mentioned cross-age face recognition method, the pixel-level semantic association structure of the first and second semantic differential feature maps of the faces of the objects to be identified is subjected to adaptive attenuation modulation based on spatial distribution guidance to obtain a semantic association weight map between spatial attenuation pixels of the semantic differential feature of the faces of the objects to be identified, including: performing fine-grained spatial decoupling of the features of the first and second semantic differential feature maps of the faces of the objects to be identified along the channel dimension to obtain a set of pixel-grained semantic differential feature vectors of the faces of the objects to be identified; any two pixel-grained semantic differential feature vectors of the faces of the objects to be identified in the set of pixel-grained semantic differential feature vectors of the faces of the objects to be identified are input into a semantic association score measurement network to obtain a semantic association weight map of the face of the objects to be identified. A set of semantic association score vectors between pixel granularities of face differential features; performing semantic association adaptive attenuation modulation on the set of semantic association score vectors between pixel granularities of face differential features of objects to be identified based on the spatial span between any two semantic association score vectors between pixel granularities of face differential features of objects to be identified in the set of semantic association score vectors between pixel granularities of face differential features of objects to be identified to obtain a set of semantic association score vectors between spatial attenuation pixel granularities of face differential features of objects to be identified; performing dimensional reconstruction and weight assignment on the set of semantic association score vectors between spatial attenuation pixel granularities of face differential features of objects to be identified to obtain a semantic association weight map between spatial attenuation pixel granularities of semantic differential features of objects to be identified.

[0017] In the above-mentioned cross-age face recognition method, any two semantic differential feature vectors of the face pixel granularity of the object to be identified in the set of semantic differential feature vectors of the face pixel granularity of the object to be identified are input into the semantic association score measurement network to obtain a set of semantic association score vectors between the pixel granularity of the face differential feature of the object to be identified, including: cascading and fusing any two semantic differential feature vectors of the face pixel granularity of the object to be identified in the set of semantic differential feature vectors of the face pixel granularity of the object to be identified, multiplying them by a weight parameter matrix, and then dot-adding them with a bias vector to obtain a set of semantic association score vectors between the pixel granularity of the face differential feature of the object to be identified.

[0018] In the above-mentioned cross-age face recognition method, based on the spatial span between any two face pixel granularity semantic differential feature vectors of the objects to be identified in the set of face pixel granularity semantic differential feature vectors of the objects to be identified, the set of semantic association score vectors between pixel granularity of face differential features of the objects to be identified is subjected to semantic association adaptive attenuation modulation to obtain a set of semantic association score vectors between pixel granularity of face differential features of the objects to be identified, including: spatially marking the spatial position of each face pixel granularity semantic differential feature vector of the objects to be identified in the set of face pixel granularity semantic differential feature vectors of the objects to be identified to obtain a set of spatial position data of face differential feature pixels of the objects to be identified; calculating the spatial position data of face differential feature pixels of the objects to be identified The Euclidean distance between the spatial position data of any two semantic differential feature vectors of the face pixel granularity of the objects to be identified in the set is used as the spatial span value of the two semantic differential feature vectors of the face pixel granularity of the objects to be identified to obtain a set of spatial span values ​​of the face differential feature pixel granularity of the objects to be identified; based on each spatial span value of the face differential feature pixel granularity of the objects to be identified in the set of spatial span values ​​of the face differential feature pixel granularity of the objects to be identified, each semantic association score vector between the pixel granularity of the face differential feature of the objects to be identified in the set of semantic association score vectors between the pixel granularity of the face differential feature of the objects to be identified is subjected to semantic association adaptive attenuation modulation to obtain a set of semantic association score vectors between the spatial attenuation pixel granularity of the face differential feature of the objects to be identified.

[0019] In the above-mentioned cross-age face recognition method, the set of semantic association score vectors between spatial attenuation pixel granularities of the face differential features of the to-be-recognized object is dimensionally reconstructed and weighted to obtain a semantic association weight map between spatial attenuation pixel granularities of the semantic differential features of the to-be-recognized object, including: arranging the set of semantic association score vectors between spatial attenuation pixel granularities of the face differential features of the to-be-recognized object into a semantic association score feature map between spatial attenuation pixel granularities of the face differential features of the to-be-recognized object, and then inputting it into a feature dimension modulation layer based on a point convolution layer and a weight allocation network based on a Sigmoid function to obtain a semantic association weight map between spatial attenuation pixel granularities of the semantic differential features of the to-be-recognized object.

[0020] In the above-mentioned cross-age face recognition method, feature modulation is performed on the first and second semantic differential feature maps of faces of the objects to be identified based on the semantic association weight map between the attenuated pixel granularity of the spatial semantic differential feature of the faces of the objects to be identified to obtain the first and second semantic differential enhanced feature maps of faces of the objects to be identified, including: calculating the position point multiplication between the semantic association weight map between the attenuated pixel granularity of the spatial semantic differential feature of the faces of the objects to be identified and the first and second semantic differential feature maps of faces of the objects to be identified to obtain the first and second semantic differential enhanced feature maps of faces of the objects to be identified.

[0021] In the above-mentioned cross-age face recognition method, based on the first and second facial semantic difference enhanced feature maps of the objects to be identified, an age difference decoding value based on facial semantics is generated, including: inputting the first and second facial semantic difference enhanced feature maps of the objects to be identified into an age difference regression prediction module based on a decoder to obtain the age difference decoding value based on facial semantics.

[0022] Compared with the prior art, the cross-age face recognition method provided by the present application uses deep learning-based image processing technology to process two cross-age face images, mines the facial structure feature representations of the two, and predicts the ages of the two based on this. At the same time, the facial structure features of the cross-age face images are subjected to difference comparison analysis, and the age difference is further predicted based on the facial difference features, so as to determine whether the face objects in the cross-age face images are the same object based on the dual-channel prediction results. Through this dual-channel comparative learning mechanism, the problem of reduced recognition accuracy of traditional face recognition technology in cross-age scenarios can be effectively solved, and the accuracy and robustness of face recognition can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0024] Figure 1 Flowchart of a cross-age face recognition method according to an embodiment of the present application.

[0025] Figure 2 Schematic diagram of data flow of the cross-age face recognition method according to an embodiment of the present application.

[0026] Figure 3 Flow chart of step S120 in the cross-age face recognition method according to an embodiment of the present application.

[0027] Figure 4 Flow chart of step S140 in the cross-age face recognition method according to an embodiment of the present application. DETAILED DESCRIPTION

[0028] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0029] In response to the problem of cross-age face recognition, this application proposes a cross-age face recognition method based on dual-channel face contrast learning, which can effectively improve the face recognition performance of cross-age scenes. Specifically, it uses image processing technology based on deep learning to process two cross-age face images, mines the facial structure feature representation of the two, and predicts the age of the two based on this. At the same time, it performs a difference comparison analysis on the facial structure features of the cross-age face images, and further predicts the age difference based on the facial difference features, so as to determine whether the face objects in the cross-age face images are the same object based on the dual-channel prediction results. Through this dual-channel contrast learning mechanism, the problem of decreased recognition accuracy of traditional face recognition technology in cross-age scenes can be effectively solved, and the accuracy and robustness of face recognition can be improved.

[0030] Figure 1 Flowchart of a cross-age face recognition method according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of the cross-age face recognition method according to an embodiment of the present application. Figure 1 and Figure 2 As shown, according to the cross-age face recognition method of the embodiment of the present application, the steps include: S110, obtaining a first face image to be recognized and a second face image to be recognized; S120, performing age prediction and difference calculation on the first face image to be recognized and the second face image to be recognized to obtain an age difference anchor value; S130, performing facial feature difference measurement on the first face image to be recognized and the second face image to be recognized to obtain a first-second face semantic difference feature map of the object to be recognized; S140, performing attention enhancement on the first-second face semantic difference feature map of the object to be recognized to obtain a first-second face semantic difference enhancement feature map of the object to be recognized; S150, based on the first-second face semantic difference enhancement feature map of the object to be recognized, generating an age difference decoding value based on face semantics; S160, in response to the difference between the age difference decoding value based on face semantics and the age difference anchor value being within a preset range, determining that the objects in the first face image to be recognized and the second face image to be recognized are the same object.

[0031] In the above-mentioned cross-age face recognition method, the step S110 obtains the first face image to be recognized and the second face image to be recognized. In the technical solution of the present application, in order to ensure that the first face image to be recognized and the second face image to be recognized have a certain age span to meet the needs of cross-age face recognition, first, a face recognition data set suitable for cross-age scenes is constructed. Specifically, an age pseudo-label is first assigned to the face recognition data set. In the face recognition data set, there are n images of the same face under the same face ID. The open source face age prediction model is used to mark the age of each image, and n labels are obtained. Finally, the average value is assigned to the face ID as the age pseudo-label of the face ID, representing the age of the face. Then, cross-age faces are generated for each image with an age pseudo-label. Different age intervals are set, and 5 age intervals are set from 0 to 100. If the pseudo-label falls in the interval of 0-20, face generation is performed in the other four intervals to simulate face images of cross-age scenes. At the end of this stage, each face ID has face images of different age intervals. Then, each face ID is divided into two subsets according to the logic of age, ensuring that the two images selected each time have a certain age span in the subsequent model training to adapt to the cross-age face recognition scenario. Based on this, the cross-age scene data set is completed. Then, the first face image to be identified and the second face image to be identified are extracted from the above cross-age scene data set, wherein the first face image to be identified and the second face image to be identified correspond to different subsets respectively, ensuring that both have a certain age span.

[0032] In the above cross-age face recognition method, the step S120 performs age prediction and difference calculation on the first face image to be recognized and the second face image to be recognized to obtain an age difference anchor value. Figure 3 FIG. 1 is a flowchart of step S120 in the cross-age face recognition method according to an embodiment of the present application. Figure 3 As shown, the step S120 includes: S121, inputting the first face image to be identified and the second face image to be identified into a face feature extractor based on a hole convolutional neural network model to obtain a first face feature map of an object to be identified and a second face feature map of an object to be identified; S122, inputting the first face feature map of an object to be identified and the second face feature map of an object to be identified into an age regression predictor based on a decoder to obtain a first age prediction value and a second age prediction value; S123, calculating the difference between the first age prediction value and the second age prediction value to obtain the age difference anchor value.

[0033] Specifically, in step S121, the first face image to be identified and the second face image to be identified are respectively input into a face feature extractor based on a dilated convolutional neural network model to obtain a face feature map of the first object to be identified and a face feature map of the second object to be identified. It should be understood that face images of different age groups have different skin textures, facial contours, facial expressions, etc. Therefore, in order to fully learn the facial features in the image and use them as a basis for cross-age face recognition, the present application adopts a dilated convolutional neural network model to construct a face feature extractor, and processes the first face image to be identified and the second face image to be identified respectively. By introducing a dilated convolution operation, the dilated convolutional neural network model can expand the receptive field while maintaining the image resolution, capture richer contextual information, and take into account the detailed features and overall facial structure features in the face image, thereby providing a more accurate feature representation for subsequent cross-age face recognition.

[0034] Specifically, in step S122, the first face feature map of the object to be identified and the second face feature map of the object to be identified are respectively input into the decoder-based age regression predictor to obtain the first age prediction value and the second age prediction value. That is, in order to improve the accuracy of cross-age face recognition, the present application adds the task of face age regression in addition to the face comparison task, and predicts the age of the first face feature map of the object to be identified and the second face feature map of the object to be identified, and uses this as prior information to provide important prior knowledge and reference basis for the subsequent face comparison task. In a specific example of the present application, the age regression predictor adopts a multi-layer perceptron structure, and after flattening the input face feature map, it is nonlinearly transformed layer by layer through the fully connected layer to decouple the age information contained therein, and then outputs the corresponding age prediction value through the output layer. In an embodiment of the present application, during the training process of the decoder-based age regression predictor, the L2 loss function is used to optimize the network to minimize the difference between the predicted age value and the true age value.

[0035] Specifically, in step S123, the difference between the first age prediction value and the second age prediction value is calculated to obtain the age difference anchor value. That is, by calculating the difference between the first age prediction value and the second age prediction value, the age difference between the two objects to be identified is quantitatively represented, thereby providing age information reference for subsequent facial feature comparison analysis. For example, if the age difference between the two images is small, but the feature map difference is large, they may belong to different facial objects; conversely, if the age difference is large, but the feature map difference is small, they may belong to different age groups of the same facial object.

[0036] In the above-mentioned cross-age face recognition method, the step S130 performs facial feature difference measurement on the first face image to be recognized and the second face image to be recognized to obtain a first-second face semantic difference feature map of the object to be recognized. In a specific example of the present application, the first face feature map of the object to be recognized and the second face feature map of the object to be recognized are input into a face semantic difference feature measurement network to obtain the first-second face semantic difference feature map of the object to be recognized, wherein the face semantic difference feature measurement network is used to calculate the position difference between the first face feature map of the object to be recognized and the second face feature map of the object to be recognized to reveal the feature difference between the two, such as the position difference and shape difference of facial organs such as eyes and nose, so as to obtain the first-second face semantic difference feature map of the object to be recognized.

[0037] In the above-mentioned cross-age face recognition method, the step S140 performs attention enhancement on the first-second semantic difference feature map of the face of the object to be identified to obtain the first-second enhanced semantic difference feature map of the face of the object to be identified. In particular, considering that in the first-second semantic difference feature map of the face of the object to be identified, different feature areas may have different importances, for example, the feature differences of key facial organs such as eyes and noses are more important for the recognition process. Therefore, in order to improve the model's ability to capture key difference features in the first-second semantic difference feature map of the face of the object to be identified, the present application proposes a feature enhancement method based on spatial context perception, which dynamically adjusts its feature weights by learning the pixel-level contextual semantic association structure of the first-second semantic difference feature map of the face of the object to be identified, and performs adaptive focusing on key features, thereby enhancing the model's sensitivity to the feature differences of key facial organs.

[0038] Figure 4 FIG. 1 is a flow chart of step S140 in the cross-age face recognition method according to an embodiment of the present application. Figure 4 As shown, the step S140 includes: S141, performing adaptive attenuation modulation based on spatial distribution guidance on the pixel-level semantic association structure of the first and second facial semantic differential feature maps of the objects to be identified to obtain a semantic association weight map between spatial attenuation pixels of the semantic differential features of the objects to be identified; S142, performing feature modulation on the first and second facial semantic differential feature maps of the objects to be identified based on the semantic association weight map between spatial attenuation pixels of the semantic differential features of the objects to be identified to obtain the first and second facial semantic differential enhanced feature maps of the objects to be identified.

[0039] Specifically, the step S141 includes: performing feature fine-grained spatial decoupling along the channel dimension on the first and second semantic differential feature maps of the faces of the objects to be identified to obtain a set of semantic differential feature vectors at the pixel granularity of the faces of the objects to be identified; inputting any two semantic differential feature vectors at the pixel granularity of the faces of the objects to be identified in the set of semantic differential feature vectors at the pixel granularity of the faces of the objects to be identified into a semantic association score measurement network to obtain a set of semantic association score vectors between pixel granularity of the differential features of the faces of the objects to be identified; and The spatial span between any two semantic differential feature vectors of the pixel granularity of the faces of the objects to be identified in the set of feature vectors performs semantic association adaptive attenuation modulation on the set of semantic association score vectors between the pixel granularity of the differential features of the faces of the objects to be identified to obtain a set of semantic association score vectors between the spatial attenuation pixel granularity of the differential features of the faces of the objects to be identified; the set of semantic association score vectors between the spatial attenuation pixel granularity of the differential features of the faces of the objects to be identified is dimensionally reconstructed and weighted to obtain a semantic association weight map between the spatial attenuation pixel granularity of the semantic differential features of the faces of the objects to be identified.

[0040] In a specific example of the present application, any two semantic differential feature vectors of the face pixel granularity of the object to be identified in the set of semantic differential feature vectors of the face pixel granularity of the object to be identified are input into the semantic association score measurement network to obtain a set of semantic association score vectors between the pixel granularity of the face differential features of the object to be identified, including: cascading and fusing any two semantic differential feature vectors of the face pixel granularity of the object to be identified in the set of semantic differential feature vectors of the face pixel granularity of the object to be identified, multiplying them by a weight parameter matrix, and then dot-adding them with a bias vector to obtain a set of semantic association score vectors between the pixel granularity of the face differential features of the object to be identified.

[0041] That is, firstly, the first and second semantic differential feature maps of the faces of the objects to be identified are subjected to fine-grained spatial decoupling along the channel dimension, and are decomposed into feature vectors of multiple pixel granularities to form a set of semantic differential feature vectors of pixel granularity of the faces of the objects to be identified, thereby enhancing the model's recognition ability of local detail features. Next, considering that local features belonging to the same facial organ usually have stronger semantic associations, the present application further measures the semantic association of any two semantic differential feature vectors of pixel granularity of the faces of the objects to be identified in the set to estimate the strength of the semantic connection between the two, and generates a set of semantic association score vectors between pixel granularity of differential features of the faces of the objects to be identified, thereby revealing the semantic association structure in the first and second semantic differential feature maps of the faces of the objects to be identified.

[0042] In a specific example of the present application, a semantic association adaptive attenuation modulation is performed on the set of semantic association score vectors between pixel granularities of face differential features of the object to be identified, including: spatially marking the semantic differential feature vectors of pixel granularity of the face of the object to be identified in the set of semantic differential feature vectors of pixel granularity of the face of the object to be identified to obtain a set of spatial position data of the face differential feature pixels of the object to be identified; calculating the Euclidean distance between the spatial position data of the semantic differential feature vectors of pixel granularity of the face of the object to be identified in the set of spatial position data of the face differential feature pixels of the object to be identified as the two The spatial span value of the semantic differential feature vector of the face pixel granularity of the object to be identified is used to obtain a set of spatial span values ​​of the differential feature pixel granularity of the object to be identified; based on each spatial span value of the differential feature pixel granularity of the object to be identified in the set of spatial span values ​​of the differential feature pixel granularity of the object to be identified, each semantic association score vector between the differential feature pixel granularity of the object to be identified in the set of semantic association score vectors between the differential feature pixel granularity of the object to be identified is subjected to semantic association adaptive attenuation modulation to obtain a set of semantic association score vectors between the spatial attenuated pixel granularity of the differential feature of the object to be identified.

[0043] In a specific example of the present application, semantic association adaptive attenuation modulation is performed on each semantic association score vector between pixel granularities of face differential features of the object to be identified in the set of semantic association score vectors between pixel granularities of face differential features of the object to be identified, including: dividing the spatial span value of the pixel granularity of the face differential features of the object to be identified by a preset scaling factor to obtain a scaling value of the spatial span of the pixel granularity of the face differential features of the object to be identified; calculating an exponential function value with e as the base and the opposite of the scaling value of the spatial span of the pixel granularity of the face differential features of the object to be identified as the exponent to obtain a semantic association spatial attenuation coefficient; calculating the semantic association score vector between pixel granularities of face differential features of the object to be identified and dividing it by the semantic association spatial attenuation coefficient to obtain the semantic association score vector between the spatial attenuated pixel granularities of face differential features of the object to be identified.

[0044] In a specific example of the present application, the set of semantic association score vectors between the spatial attenuation pixel granularities of the face differential features of the object to be identified is dimensionally reconstructed and weighted, including: after arranging the set of semantic association score vectors between the spatial attenuation pixel granularities of the face differential features of the object to be identified into a semantic association score feature map between the spatial attenuation pixel granularities of the face differential features of the object to be identified, it is input into a feature dimension modulation layer based on a point convolution layer and a weight allocation network based on a Sigmoid function to obtain a semantic association weight map between the spatial attenuation pixel granularities of the semantic differential features of the face of the object to be identified.

[0045] Here, considering that in an image, similar pixels often have more similar semantic information, the present application further annotates the spatial position of the pixel-granular semantic differential feature vectors of the faces of the objects to be identified, adds the exact position information in the original feature map, and calculates the spatial span between any two pixel-granular semantic differential feature vectors of the faces of the objects to be identified based on this, thereby revealing the relative position relationship between the local features of each pixel-granularity, so as to filter out the interference items that are similar on the surface but not actually related in the subsequent feature modulation process. Then, based on the spatial span value between the pixel-granular semantic differential feature vectors of the faces of any two objects to be identified, the semantic association score vector between the two is adaptively attenuated and modulated. When two pixel-granular local features show strong semantic relevance and are close to each other, the importance of both will be maximized; on the contrary, if the two have a high semantic similarity but are far apart, their influence will be appropriately weakened, thereby accurately reflecting the most noteworthy feature parts in the first-second semantic differential feature map of the faces of the objects to be identified. Then, the set of semantic association score vectors between pixel granularities of face differential features to be identified after spatial attenuation modulation is arranged into a feature map, and then the feature dimension is modulated and weighted by point convolution and Sigmoid function to generate the corresponding semantic association weight map between pixel granularities of face semantic differential features to be identified.

[0046] Specifically, the step S142 includes: calculating the semantic association weight map between the spatial attenuation pixel granularity of the semantic difference feature of the face of the object to be identified and multiplying the first and second semantic difference feature maps of the face of the object to be identified by the position point to obtain the first and second semantic difference enhanced feature maps of the face of the object to be identified. That is, the first and second semantic difference feature maps of the face of the object to be identified are weighted modulated by element-by-element multiplication operation, so as to selectively amplify important information and suppress irrelevant background information or non-critical feature areas, highlight the expression of key facial difference features, and thus obtain the first and second semantic difference enhanced feature maps of the face of the object to be identified.

[0047] Accordingly, the step S140 includes: processing the first and second to-be-recognized object face semantic difference feature maps using the following feature enhancement formula to obtain the first and second to-be-recognized object face semantic difference enhanced feature maps, wherein the feature enhancement formula is:

[0048]

[0049]

[0050]

[0051]

[0052]

[0053]

[0054]

[0055]

[0056]

[0057]

[0058]

[0059]

[0060]

[0061] in, is the semantic difference feature map of the face of the first and second objects to be identified, and ,in, , and are respectively the number of channels, height and width of the semantic difference feature maps of the first and second objects to be identified, represents the feature decoupling operation, is a set of pixel-level semantic differential feature vectors of the face of the object to be identified. is the pixel-level semantic differential feature vector of the face of the object to be identified, and is the length of the pixel-granular semantic differential feature vector of the face of the object to be identified, is the spatial position data of the pixel-granular semantic differential feature vector of the face of the object to be identified, where: , is the spatial position coordinate of the pixel-granular semantic differential feature vector of the face of the object to be identified, Represents the set of spatial position data of the face differential feature pixels of the object to be identified. and Any two face pixel granularity semantic difference feature vectors of the to-be-recognized object in the set of face pixel granularity semantic difference feature vectors of the to-be-recognized object are represented, and Represent the weight parameter matrix and bias vector respectively, Represents the semantic association score vector between pixel granularity of face differential features of the object to be identified, is a set of semantic association score vectors between pixel granularities of the face differential features of the object to be identified, is the spatial span value between the pixel-granular semantic difference feature vectors of any two faces of the objects to be identified, is the set of pixel granularity spatial span values ​​of the face differential feature of the object to be identified, is the preset scaling factor, is the semantic association score vector between the attenuated pixel granularity of the face differential feature space of the object to be identified, is the set of semantic association score vectors between attenuated pixel granularities of the face differential feature space of the object to be identified, Represents the semantic association score feature map between the pixel granularity of the face difference space attenuation of the object to be identified, represents the point convolution operation, 𝐴' is the semantic association score feature map between the pixel granularity of the face difference space attenuation of the object to be identified after feature dimension modulation, is the Sigmoid function, Represents the semantic association weight map between the attenuated pixel granularity of the semantic differential feature space of the face of the object to be identified, It means point multiplication by position. Represents the semantic difference enhanced feature map of the first and second objects to be identified.

[0062] In the above-mentioned cross-age face recognition method, the step S150 generates an age difference decoding value based on face semantics based on the first and second face semantic difference enhancement feature maps of the objects to be recognized. In a specific example of the present application, the step S150 includes: inputting the first and second face semantic difference enhancement feature maps of the objects to be recognized into the age difference regression prediction module based on the decoder to obtain the age difference decoding value based on face semantics. In the specific scheme of the present application, the age difference regression prediction module based on the decoder also adopts a multi-layer perceptron structure, which performs feature analysis on the first and second face semantic difference enhancement feature maps of the objects to be recognized through layer-by-layer nonlinear transformation, and maps the face difference information contained therein into an age difference decoding value. The feature difference between the two objects to be recognized is quantitatively represented by the age difference decoding prediction based on the face semantic difference feature. In an embodiment of the present application, in the training process of the age difference regression prediction module based on the decoder, the InfoNCE (Information Noise Contrastive Estimation) loss function is used for network optimization.

[0063] Preferably, inputting the first-second to-be-recognized object face semantic differential enhanced feature map into the decoder-based age difference regression prediction module to obtain the age difference decoding value based on face semantics includes: determining the number of zero eigenvalues ​​in the first-second to-be-recognized object face semantic differential enhanced feature map, and respectively calculating the reciprocal of the logarithm of the number of zero eigenvalues ​​with base two and the exponential value of the reciprocal of the number of zero eigenvalues ​​with base natural constant to obtain the first to-be-recognized object face semantic differential representation value and the second to-be-recognized object face semantic differential representation value; expanding the first-second to-be-recognized object face semantic differential enhanced feature map into the first-second to-be-recognized object face semantic differential enhanced feature vector; calculating the power function of each eigenvalue of the first-second to-be-recognized object face semantic differential enhanced feature vector with the reciprocal of the number of zero eigenvalues ​​as the exponent to obtain the first-second to-be-recognized object face semantic differential structural state vector; The first-second to-be-recognized object face semantic difference structure state vector is dot-multiplied with the first-second to-be-recognized object face semantic difference representation value to obtain the first-second to-be-recognized object face semantic difference information representation vector; the autocorrelation matrix of the first-second to-be-recognized object face semantic difference structure state vector is dot-multiplied with the second first-second to-be-recognized object face semantic difference representation value and weight hyperparameter respectively to obtain the first-second to-be-recognized object face semantic difference regression understanding matrix; the first-second to-be-recognized object face semantic difference information representation vector is matrix-multiplied with the first-second to-be-recognized object face semantic difference regression understanding matrix to obtain the optimized first-second to-be-recognized object face semantic difference enhanced feature vector; the optimized first-second to-be-recognized object face semantic difference enhanced feature vector is input into the decoder-based age difference regression prediction module to obtain the age difference decoding value based on face semantics.

[0064] Here, the semantic difference representation value of the face of the first object to be identified and the semantic difference representation value of the face of the second object to be identified are expressed as:

[0065]

[0066]

[0067] in, is the number of zero eigenvalues ​​in the semantic differential enhanced feature map of the first and second objects to be identified, is the semantic difference representation value of the face of the first object to be identified, the semantic difference representation value of the face of the second object to be identified, represents the logarithmic function with base 2, represents the natural exponential function.

[0068] And the optimized first-second face semantic difference enhanced feature vector It is expressed as:

[0069]

[0070]

[0071]

[0072] Among them, the semantic difference enhanced feature vector of the first and second objects to be identified is a row vector, represents a power function in which each eigenvalue of the eigenvector is calculated with the inverse of the number of zero eigenvalues ​​as its exponent, represents the transpose of a vector, represents the matrix multiplication operation, is the weight hyperparameter, represents the semantic difference regression understanding matrix of the first and second objects to be recognized. A vector representing the semantic difference information of the first and second objects to be identified.

[0073] Here, when the first face feature map of the object to be identified and the second face feature map of the object to be identified respectively represent the face image semantic coding features of the first face image to be identified and the second face image to be identified, when performing face semantic difference feature measurement and feature attention enhancement based on spatial modulation, the cross-domain face semantic difference features of the first-second face semantic difference enhancement feature maps for the objects to be identified will have a long-range class regression understanding representation imbalance due to spatial flexible modulation, thereby affecting the accuracy of the decoding results.

[0074] Based on this, the present application realizes effective long-range modeling of the first-second object to be identified face semantic differential enhancement feature map under field linear complexity by modeling the vector field zero dimension of the first-second object to be identified face semantic differential enhancement feature vector obtained after the first-second object to be identified face semantic differential enhancement feature map is expanded as the global field redundant dependency, so as to perform structured correlation self-distillation between information representation and class understanding in a selective state space based on feature distribution in the feature regression process of the first-second object to be identified face semantic differential enhancement feature map, thereby maintaining the overall information complexity of the feature distribution of the first-second object to be identified face semantic differential enhancement feature map while promoting the long-range dynamic class understanding balance of its feature convergence, so as to improve the decoding understanding consistency between the decoding regression and feature extraction processes, and improve the accuracy of the age difference decoding value based on facial semantics obtained by inputting the first-second object to be identified face semantic differential enhancement feature map into the decoder-based age difference regression prediction module.

[0075] In the above-mentioned cross-age face recognition method, in step S160, in response to the difference between the age difference decoding value based on facial semantics and the age difference anchor value being within a preset range, it is determined that the object in the first face image to be identified and the second face image to be identified is the same object. That is, based on the comprehensive judgment of facial feature similarity and age factor verification, the final confirmation of face recognition is completed. For example, when the difference is less than the set threshold, it means that the degree of facial difference between the two images conforms to the natural variation range of the same face object in different age groups, so they can be confirmed as the same person. On the contrary, if the difference exceeds the threshold, it means that the degree of facial difference between the two images exceeds the natural variation range of the same face object in different age groups, so they can be determined as different individuals.

[0076] In summary, the cross-age face recognition method according to the embodiment of the present application is explained, which uses deep learning-based image processing technology to process two cross-age face images, mines the facial structure feature representations of the two, and predicts the ages of the two based on this. At the same time, the facial structure features of the cross-age face images are subjected to difference comparison analysis, and the age difference is further predicted based on the facial difference features, so as to determine whether the face objects in the cross-age face images are the same object based on the dual-channel prediction results. Through this dual-channel comparative learning mechanism, the problem of reduced recognition accuracy of traditional face recognition technology in cross-age scenarios can be effectively solved, and the accuracy and robustness of face recognition can be improved.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A cross-age face recognition method, characterized in that: include: Acquire a first face image to be recognized and a second face image to be recognized; Performing age prediction and difference calculation on the first face image to be recognized and the second face image to be recognized to obtain an age difference anchor value; Performing facial feature difference measurement on the first to-be-recognized face image and the second to-be-recognized face image to obtain a first-to-second to-be-recognized face semantic difference feature map; Performing attention enhancement on the first and second face semantic difference feature maps of the objects to be identified to obtain the first and second face semantic difference enhanced feature maps of the objects to be identified; Based on the semantic difference enhanced feature map of the first and second objects to be identified, generating an age difference decoding value based on facial semantics; In response to a difference between the age difference decoding value based on facial semantics and the age difference anchor value being within a preset range, determining that the objects in the first to-be-recognized face image and the second to-be-recognized face image are the same object; Among them, the attention is enhanced for the first and second semantic differential feature maps of faces of objects to be identified, including: performing adaptive attenuation modulation based on spatial distribution guidance on the pixel-level semantic association structure of the first and second semantic differential feature maps of faces of objects to be identified to obtain a semantic association weight map between spatial attenuation pixels of the semantic differential feature of faces of objects to be identified; based on the semantic association weight map between spatial attenuation pixels of the semantic differential feature of faces of objects to be identified, feature modulation is performed on the first and second semantic differential feature maps of faces of objects to be identified to obtain the first and second semantic differential enhanced feature maps of faces of objects to be identified.

2. The cross-age face recognition method according to claim 1, characterized in that: Performing age prediction and difference calculation on the first to-be-recognized face image and the second to-be-recognized face image to obtain an age difference anchor value, including: Inputting the first face image to be recognized and the second face image to be recognized into a face feature extractor based on a dilated convolutional neural network model to obtain a first face feature map of an object to be recognized and a second face feature map of an object to be recognized; Inputting the first face feature map of the object to be identified and the second face feature map of the object to be identified into the decoder-based age regression predictor to obtain a first age prediction value and a second age prediction value; The difference between the first age prediction value and the second age prediction value is calculated to obtain the age difference anchor value.

3. The cross-age face recognition method according to claim 2, characterized in that: Measuring the difference of facial features of the first to-be-recognized facial image and the second to-be-recognized facial image to obtain a semantic difference feature map of the first to-be-recognized object's facial features, including: The first face feature map of the object to be identified and the second face feature map of the object to be identified are input into a face semantic difference feature measurement network to obtain the first-second face semantic difference feature map of the object to be identified, wherein the face semantic difference feature measurement network is used to calculate the positional difference between the first face feature map of the object to be identified and the second face feature map of the object to be identified to obtain the first-second face semantic difference feature map of the object to be identified.

4. The cross-age face recognition method according to claim 3, characterized in that: The pixel-level semantic association structure of the first and second face semantic difference feature maps of the object to be identified is subjected to adaptive attenuation modulation based on spatial distribution guidance to obtain a semantic association weight map between spatial attenuation pixels of the face semantic difference feature of the object to be identified, including: Performing feature fine-grained spatial decoupling along the channel dimension on the first and second semantic differential feature maps of the faces of the objects to be identified to obtain a set of semantic differential feature vectors of pixel granularity of the faces of the objects to be identified; Inputting any two semantic difference feature vectors of face pixel granularity of the to-be-identified object from the set of semantic difference feature vectors of face pixel granularity of the to-be-identified object into the semantic association score measurement network to obtain a set of semantic association score vectors between the pixel granularity of face difference feature of the to-be-identified object; Based on the spatial span between any two semantic differential feature vectors of pixel granularity of faces of the objects to be identified in the set of semantic differential feature vectors of pixel granularity of faces of the objects to be identified, the set of semantic association score vectors between pixel granularity of face differential features of the objects to be identified is subjected to semantic association adaptive attenuation modulation to obtain a set of semantic association score vectors between pixel granularity of face differential features of the objects to be identified; The set of semantic association score vectors of the spatial attenuation pixel granularity of the face differential feature of the object to be identified is dimensionally reconstructed and weighted to obtain a semantic association weight map of the spatial attenuation pixel granularity of the semantic differential feature of the face of the object to be identified.

5. The cross-age face recognition method according to claim 4, characterized in that: Inputting any two semantic difference feature vectors of the face pixel granularity of the object to be identified from the set of semantic difference feature vectors of the face pixel granularity of the object to be identified into the semantic association score measurement network to obtain a set of semantic association score vectors between the face difference feature pixel granularity of the object to be identified, including: Any two semantic differential feature vectors of the pixel granularity of the faces of the objects to be identified in the set of semantic differential feature vectors of the pixel granularity of the faces of the objects to be identified are cascaded and fused, multiplied by a weight parameter matrix, and then dotted with a bias vector to obtain a set of semantic association score vectors between the pixel granularity of the differential features of the objects to be identified.

6. The cross-age face recognition method according to claim 5, characterized in that: Based on the spatial span between any two semantic differential feature vectors of the pixel granularity of the face of the object to be identified in the set of the semantic differential feature vectors of the pixel granularity of the face of the object to be identified, the set of semantic association score vectors between pixel granularity of the face differential feature of the object to be identified is subjected to semantic association adaptive attenuation modulation to obtain a set of semantic association score vectors between pixel granularity of the face differential feature of the object to be identified, including: Performing spatial position annotation on each of the face pixel granularity semantic differential feature vectors of the to-be-identified object in the set of face pixel granularity semantic differential feature vectors of the to-be-identified object to obtain a set of face differential feature pixel spatial position data of the to-be-identified object; Calculating the Euclidean distance between the spatial position data of any two face pixel granularity semantic differential feature vectors of the objects to be identified in the set of face differential feature pixel spatial position data of the objects to be identified as the spatial span value of the two face pixel granularity semantic differential feature vectors of the objects to be identified to obtain a set of face differential feature pixel granularity spatial span values ​​of the objects to be identified; Based on each face differential feature pixel granularity spatial span value of the object to be identified in the set of face differential feature pixel granularity spatial span values ​​of the object to be identified, semantic association adaptive attenuation modulation is performed on each face differential feature pixel granularity semantic association score vector of the object to be identified in the set of face differential feature pixel granularity semantic association score vectors to obtain the set of face differential feature spatial attenuated pixel granularity semantic association score vectors of the object to be identified.

7. The cross-age face recognition method according to claim 6, characterized in that: The set of semantic association score vectors of the space attenuation pixel granularity of the face difference feature of the object to be identified is reconstructed in dimension and weighted to obtain a semantic association weight map of the space attenuation pixel granularity of the face semantic difference feature of the object to be identified, including: After arranging the set of semantic association score vectors between spatial attenuation pixel granularities of the face differential features of the object to be identified into a semantic association score feature map between spatial attenuation pixel granularities of the face differential features of the object to be identified, it is input into a feature dimension modulation layer based on a point convolution layer and a weight allocation network based on a Sigmoid function to obtain a semantic association weight map between spatial attenuation pixel granularities of the semantic differential features of the object to be identified.

8. The cross-age face recognition method according to claim 7, characterized in that: The method comprises: performing feature modulation on the first and second to-be-recognized object face semantic difference feature maps based on the semantic association weight map between the attenuated pixel granularity of the face semantic difference feature space of the to-be-recognized object to obtain the first and second to-be-recognized object face semantic difference enhanced feature maps, including: Calculate the semantic association weight map between the spatial attenuation pixel granularity of the semantic difference feature of the face of the object to be identified and multiply the first and second semantic difference feature maps of the faces of the objects to be identified by the position points to obtain the first and second semantic difference enhanced feature maps of the faces of the objects to be identified.

9. The cross-age face recognition method according to claim 8, characterized in that: Based on the first and second face semantic difference enhanced feature maps of the objects to be identified, generating an age difference decoding value based on face semantics, including: The first and second facial semantic difference enhanced feature maps of the objects to be identified are input into the decoder-based age difference regression prediction module to obtain the facial semantic-based age difference decoding value.

Images