Research On Animal Facial Recognition Methods

Identification of individual animals is foundational to smart livestock management.Facial recognition technology offers significant advantages over other identification methods,including safety,convenience,and non-invasiveness.However,current animal facial recognition technology faces challenges such as data complexity and limitations,deficiencies in feature extraction performance of deep learning network models,and constraints in detecting small targets.Additionally,the substantial differences between animal and human facial features,along with the difficulties in recognizing non-frontal animal faces,further complicate the identification task.To address these challenges,this paper first proposes an adaptive activation function aimed at enhancing the performance of deep learning network models.Subsequently,a robust and efficient animal facial detection algorithm specifically designed for small targets has been developed to obtain high-quality animal facial images.Furthermore,by integrating autoencoder networks with transformation networks,a facial feature domain transformation network model has been established to facilitate the direct application of human facial recognition technology by converting animal faces into virtual human faces.Finally,by integrating autoencoder networks,high-resolution reconstruction networks,and generative adversarial networks,a high-resolution frontal face generation model has been developed to improve the accuracy of animal facial recognition.The main research contents of this paper are as follows:(1)To address the challenges of data complexity and limitations in the field of animal face recognition,as well as the insufficiency of deep learning network models in extracting subtle features,this paper explores a new activation function design method based on the theory of squashing functions through an in-depth analysis of non-saturating activation functions.Furthermore,by hardening the squashing functions,the concept of ”segmented squashing functions” is proposed,and innovatively,the Quadratic Linear Unit(Qu LU)activation function is introduced.By setting the parameters α and β of the Qu LU as trainable parameters,the Adaptive Quadratic Linear Unit(AQu LU)activation function is proposed.This activation function effectively overcomes issues of gradient vanishing,gradient explosion,and the dying Re LU problem,with significant optimization in computational cost.In image classification tasks,it achieved up to 21.86% and 4.49% performance improvement on the CIFAR-10 and CIFAR-100 datasets,respectively,compared to the Re LU function,demonstrating its significant advantage in enhancing network performance.(2)To tackle the accuracy constraints of detection algorithms due to small targets and real-time requirements,an efficient Automatic Matching Target Network is proposed,employing Efficient Net and Recursive Feature Pyramid(RPN)as the feature extraction and fusion networks,respectively.Through innovative grid division strategies and traversal pooling layer design,the precision of target detection is effectively improved.The analysis shows that the method of three concentric rectangle boxes and dynamically adjusting the number of grids optimizes the matching process.Additionally,by optimizing the balance of positive and negative samples,the model performance is further enhanced,with tests on the MS COCO dataset showing that the network achieves an APs of 29.5% and an AP of 46% for small target detection accuracy.For animal face recognition applications,compared to a cow face detector based on Mask-RCNN,this network maintains similar detection accuracy with a faster processing speed.(3)To address the significant difference in facial features between animals and humans,a facial feature domain transformation network model is developed to convert the frontal facial feature domain of animals into that of humans,while ensuring the uniqueness of the transformed features is preserved.The network structure includes an autoencoder and a transcoder network,where the autoencoder is responsible for converting animal and human facial images into three 256-dimensional low-dimensional feature vectors,capturing their respective facial feature domains.The transcoder network further establishes a mapping between the two feature domains,achieving the conversion of animal faces to virtual human faces.Tests on pet(cat,dog)and livestock(horse,cow)facial images show that the network can accurately convert cross-species facial images,maintaining individual differences and intra-individual consistency.Experimental results using four face recognition models for animal face recognition show an accuracy rate exceeding 94.45%,with the VGG-Face model reaching an accuracy rate of 99.75%.(4)To solve the challenge of recognizing low-resolution,non-frontal animal faces,a High-Resolution Reconstruction and Correction Network(HRRCN)is proposed,which converts low-resolution and non-frontal animal facial images into frontal faces while significantly enhancing resolution.The network contains a frontalization autoencoder subnetwork to adjust the pose,a resolution enhancement sub-network to improve clarity,and a discriminator network to enhance image realism.Optimized through a comprehensive loss function,the network ensures the identity consistency and visual realism of the images.Experimental results demonstrate the HRRCN’s exceptional performance in restoring high-definition frontal face images,and the restored frontal images are compatible with existing face recognition models,achieving accuracies of 98.75% and 96.25%on horse and cow facial datasets,respectively.