CN-121214492-B - Multi-mode recognition method for nose patterns and faces of dogs

CN121214492BCN 121214492 BCN121214492 BCN 121214492BCN-121214492-B

Abstract

A multi-mode recognition method for nose lines and faces of dogs belongs to the field of computer vision and mode recognition. In order to solve the problem that the traditional method only uses nose patterns or only uses faces and has high false recognition/false recognition rate, the method firstly sends a canine face image into two branches for processing, in the canine nose pattern processing branches, firstly adopts a network model to process the image to obtain a standardized nose pattern image, obtains nose pattern embedded vectors based on the nose pattern image, searches in a nose pattern search library, and calculates the nose pattern distance by using cosine similarity In the canine face processing branch, canine face embedded vectors are obtained based on canine face images, cosine distance matching is carried out on the canine face embedded vectors and the canine face candidate IDs corresponding to the nasal vein candidate IDs, and the canine face distance is calculated by utilizing cosine similarity Based on And And obtaining a fusion score, further obtaining the Top-N with the highest score and the confidence coefficient thereof, and realizing canine identification.

Inventors

DU JIAHUI
DAI CHANGQING
ZHANG ZHENG
LIU XIN

Assignees

喵汪鲜研(北京)科技有限公司
包头市阿宝宠物连锁经营有限公司

Dates

Publication Date: 20260512
Application Date: 20251017

Claims (6)

1. A method for multi-modal identification of canine nose print and facial features comprising: firstly, sending a canine face image into two branches for processing, wherein the two branches comprise a canine face processing branch and a canine nose line processing branch; The dog nose pattern processing branch comprises the steps of adopting a second network model to process a dog face image, outputting a dog nose candidate frame, performing non-maximum suppression on the candidate frame to remove an overlapped frame, obtaining a nose pattern frame with highest confidence degree, further obtaining a standardized nose pattern image, then sending the nose pattern image into a third network model which is the same as the first network model in structure and different in network parameters to obtain a nose pattern embedded vector, performing normalization processing on the nose pattern embedded vector, performing retrieval in a nose pattern retrieval library, and calculating the nose pattern distance by using cosine similarity And according to As confidence of nose print According to the nose patterns, the nose pattern candidate IDs are arranged in ascending distance sequence, and the corresponding nose pattern confidence coefficient is obtained; The canine face processing branch comprises obtaining a canine face embedded vector through a first network model based on a canine face image, normalizing the canine face embedded vector, performing cosine distance matching with Top-K canine face candidate IDs, and calculating the canine face distance by using cosine similarity And according to As confidence of canine face The Top-K dog face candidate IDs are based on the Top-K nose line candidate IDs, and the corresponding dog face candidate IDs are extracted from the dog face search library; Then based on the nose line distance Distance from face of dog Obtaining a fusion score , wherein, And Is the fusion weight and is based on the fusion score Finally, the Top-N with the highest fusion confidence score and the confidence coefficient thereof are output, thereby realizing the identification of dogs; in the third network model training process, the loss function used comprises highlight invariable consistency loss The said The following are provided: , , Wherein, the For the third network model to be a model of the network, Is a nose pattern image The nose print embedding vector obtained by the third network model, To remove highlight nose pattern image Nose pattern embedded vector obtained through third network model and highlight nose pattern removing image Is a nose pattern image An image obtained after the highlight is removed; And As an intermediate variable, the number of the variables, Representing a binary norm; representing the calculated cosine distance.
2. The method for multimodal recognition of canine nasal patterns and facial patterns according to claim 1 wherein the second network model is a YOLOv network model in the canine nasal pattern processing branch at the canine face.
3. The method for multi-modal identification of canine nasal patterns and facial features of claim 1 wherein L2 normalization is used for normalization of the nasal pattern embedded vectors in the canine nasal pattern processing branches.
4. The method for multi-modal identification of canine nose print and facial features of claim 1 wherein in the canine face processing branch, the first network model uses a Resnext-101 model based on ArcFace modification.
5. The method for multi-modal identification of canine nose print and facial features of claim 1 wherein the normalization of the canine face embedded vector is performed using L2 normalization.
6. The method of claim 1, wherein the loss function used in the training of the first network model comprises a loss of consistency of key point-shape The said Obtained by the following steps: the key points are coded into shape description with unchanged scale/rotation after affine normalization, and are aligned to an embedding space through projection, so that consistency of facial embedding and geometric shape is restrained; Parallel lightweight keypoint regression header prediction on backbone network features To (3) pair Affine/scale normalization is performed: , Wherein mu represents the mean value, s represents the standard deviation, and l represents the number of regression heads of the key points; Representing a binary norm; Will be Encoding into shape descriptions By projection at the same time Embedding the dog face obtained by the first network model Aligning to the same dimension, and then calculating the loss: 。

Description

Multi-mode recognition method for nose patterns and faces of dogs Technical Field The invention belongs to the field of computer vision and pattern recognition, and particularly relates to a canine identification method. Background In recent years, the feeding, breeding and management of pet dogs and working dogs have huge market scale, so that the dogs are required to be identified in various aspects of urban dog raising management and registration, dog identification authentication and recovery, re-filing of a housing house/breeding ground, immunization and quarantine traceability, insurance claim anti-fraud, port/community rapid verification, scientific research, blood system management and the like. Although computer vision and pattern recognition techniques are now rapidly evolving, they are being used in a variety of applications. However, the current individual identification technology is generally aimed at the field with individual identification requirements, such as face recognition and the like, and is generally aimed at class-oriented identification of animals, such as which kind of animals are in an identification picture during identification, and even if the requirements of more specific identification are met, the technology is generally aimed at variety identification, such as which kind of dogs are golden fur or tady, and the technology is generally not capable of specifically identifying who a dog is, so that the technology for identifying the individual animal is not perfect, wherein the actual data is a factor which severely restricts the individual animal identification to be realized by a few samples of individual dogs. In addition, although other individual identification techniques can be applied to individual identification of dogs, there are problems that animals are generally not kept clean and tidy all the time, even if an animal owner, especially a dog owner, keeps the animals clean based on personal habits, but the cleanliness is only the most of cases, once the dog is taken out, the dog cleaning state cannot be ensured, and the characteristic mining and capturing difficulty for distinguishing the individual animals is also higher for the animals, so that the problem that the mistakes and/or the mistakes are higher when the mistakes and the mistakes are caused by masking, the stains, the side faces and the illumination are caused by the tradition of only nose lines or only faces. For example, the dog nose is a small target, so that the segmentation and recognition difficulty is relatively high, the nose area target is small, the boundary is thin and is often shielded by mud points/water stains, the recognition effect is seriously affected, and the robustness of a recognition model is poor. Disclosure of Invention The invention aims to solve the problem that the error recognition/miss recognition rate is high when only nose lines or only faces are used in the prior art. A method for multi-modal identification of canine nose prints and faces, comprising: firstly, sending a canine face image into two branches for processing, wherein the two branches comprise a canine face processing branch and a canine nose line processing branch; The dog nose pattern processing branch comprises the steps of adopting a second network model to process a dog face image, outputting a dog nose candidate frame, performing non-maximum suppression on the candidate frame to remove an overlapped frame, obtaining a nose pattern frame with highest confidence degree, further obtaining a standardized nose pattern image, then sending the nose pattern image into a third network model which is the same as the first network model in structure and different in network parameters to obtain a nose pattern embedded vector, performing normalization processing on the nose pattern embedded vector, performing retrieval in a nose pattern retrieval library, and calculating the nose pattern distance by using cosine similarity And according to 1-As confidence of nose printAccording to the nose patterns, the nose pattern candidate IDs are arranged in ascending distance sequence, and the corresponding nose pattern confidence coefficient is obtained; The canine face processing branch comprises obtaining a canine face embedded vector through a first network model based on a canine face image, normalizing the canine face embedded vector, performing cosine distance matching with Top-K canine face candidate IDs, and calculating the canine face distance by using cosine similarity And according to 1-As confidence of canine faceThe Top-K dog face candidate IDs are based on the Top-K nose line candidate IDs, and the corresponding dog face candidate IDs are extracted from the dog face search library; Then based on the nose line distance Distance from face of dogObtaining a fusion scoreAnd according to the integrated scoreAnd finally outputting the Top-N with the highest score and the confidence coefficient thereof, thereby realizing th