Research On Face Recognition Algorithm In Unrestricted Scenarios And Its System Implementation

With the development of modern society and technology,various face recognition applications have been widely used in a variety of scenarios,such as large public systems including epidemic prevention and control system,security and monitoring system,and financial payment system,personal devices including PC,cell phone,laptop,UAV,smart home appliances and so on,which have greatly facilitated our daily life and highly improved the efficiency of production and work.Face recognition technologies are becoming more and more indispensable and changing people’s lifestyle.The existing face recognition methods have achieved great success under restricted conditions(the ideal environment),even exceeding the performance of human eyes.However,under unrestricted scene(the wild environment),the input face images are usually degraded by external interference such as light,noise,blur and low resolution,and internal differences like pose,expression,age,skin color,occlusion and makeup,which weakens the performance of current algorithms.Therefore,it is important and necessary to research the face recognition solution under unrestricted conditions,which can not only improve the social production efficiency,but also further promote face recognition applications to more complex and extensive scenarios,so as to create more social value.To further improve the accuracy and robustness of existing face recognition systems under unrestricted scenes,this thesis mainly studies four key stages of recognition system,including face detection,face image super-resolution,face feature extraction and recognition.The main research contents and contributions are listed as follows:(1)To solve the problem of low training efficiency of face detection model and performance degradation in multi-scale and multi-pose scenarios,we proposed an adaptive anchor matching strategy(AAMS)for face detection,which deprecates the fixed matching mechanism used by previous methods.Specifically,we propose a robust anchor setting determined by the statistical characteristic of the training data,which improves the utilization of small faces.We also selectively assign proper targets to anchors of different feature levels by using adaptive matching thresholds,which promotes the accuracy and robustness of multi-scale and multi-pose face detection.Moreover,we adaptively adjust the matching threshold for each anchor according to the model’s real-time data fitting degree,enabling more flexible and precise control of the quality and quantity of the matched samples,which finally boosts the anchor matching efficiency and model convergence.Different from prior methods adopting heuristic anchor setting,centralized target assignment and fixed matching threshold,we thoroughly carried out the idea of adaptive and dynamic adjustment in the design of anchor matching strategy.Extensive experiments on popular benchmarks reveal that the proposed approach has significant improvements on anchorbased models and outperforms the recent state-of-the-arts methods in terms of both accuracy and generalization under unrestricted scenes.(2)The previous methods mainly focus on improving the visual effects of the images while retaining a challenge of restoring the identity information from low-resolution faces,which leads to huge performance degradation for low-and cross-resolution face recognition.To address this issue,we propose a facial mask attention network(FMANet)for identity-aware face superresolution to restore the identity information and image quality from low-resolution inputs suffering from multiple disturbance,which ensures that the faces detected in the previous stage are effectively recognized.Specifically,we propose the facial mask map containing vast face structure information as prior information to guide the model’s restoration process to focus on the key areas with rich identity features,which helps to generate realistic and natural facial textures.The proposed FMANet provides an end-to-end solution of face super-resolution,which effectively recovers both the identity information and the image quality.We also present Mask Pix loss and the improved perceptual loss to supervise the model learning identity features in key regions to produce highly discriminative face images that are close to real distributions.Different from the previous methods only focusing on improving the image quality metrics at the pixel level,the proposed method additionally considers the restoration of identity information at the feature level.Experimental results show that our approach can not only generate high-quality face images that are de-blurred and noiseless,but also significantly restore the identity information to a large extent,laying a solid foundation for the subsequent stages of face feature extraction and recognition algorithms.(3)To overcome the weakness of the supervising force of single action form produced by current loss functions,which are difficult to effectively constrain the feature distribution,we propose Origin loss integrating multi-action forms supervising forces from Euclidean and angular spaces to extract more discriminative and robust face features,aiming at enhancing the intra-class compactness and inter-class separation of the feature distribution.Specifically,the proposed supervising force is composed of three parts: 1)the point-to-point intra-class compact force to decrease the intra-class variance,2)the tangential inter-class force to enlarge the interclass margin,and 3)the radial origin repulsive force to push the samples near the origin away and make them easier to classify.In addition,we also analyze the action form and correctness of the proposed supervising force in mathematics and geometry.Unlike the existing studies only generating supervising force of single action form,our method can produce a powerful binding force of multiple action forms to flexibly drive the samples to move in more freedom degree.The experimental results show that the proposed approach supervises the model learning a good feature distribution,and achieves better performance than existing algorithms on various popular face recognition datasets.(4)To tackle the problem of slow training speed,low convergence level and poor robustness when applying the proposed algorithm in the last part(also existing in many prior research)to large-scale training datasets,we propose an adaptive weight modulation method for training large-scale face recognition models guided by sample Hardness,dynamically providing the model with samples of proper difficulty,i.e.,Hardness,to enhance the model’s learning ability for complex facial features.Specifically,we propose a training status estimator to evaluate the model’s data fitting degree in real time by drawing lessons from hard example mining and curriculum learning,which is employed to guide the adaptive weight modulation.We also present the definition of Hardness for each sample to describe its current difficulty to be learned.According to Hardness,we adaptively assign weights for training samples to emphasize or suppress them in different training stages.In early training process,easier samples receive larger weight and models focus on learning elementary features to improve the classification ability apace.In later training process,we emphasize the harder samples to guide models to learn more complex face features and further boost the training convergence.Different from the previous work,the proposed method take both the macrometric(e.g.,the model’s training status)and micrometric(e.g.,the sample position in feature space)into account,which is benefit for precisely and comprehensively measuring the sample’s difficulty and conducting weight modulation.Extensive test results indicate that the proposed approach not only eliminates the aforementioned issues without affecting the inference speed,but also improves the performance of the algorithm in the last chapter under unrestricted conditions,i.e.,large-scale databases.(5)To unfreeze the limitations of applying the existing face recognition systems into unrestricted scenarios,we integrate the aforementioned sub-algorithms of different face recognition stages,and implement an intelligent monitoring system for face recognition based on RK3399 Pro Ds platform.In specific,the proposed system has powerful functions such as face detection,face image super-resolution,face feature extraction and recognition,visual human-computer interaction and online database management.We have deployed and tested our system in serval different wild environments to evaluate its performance in unrestricted conditions.Experimental results demonstrate that our system not only has excellent face recognition performance in unrestricted scenes,but also has good real-time and stability.This thesis combines theoretical research with practical application,which not only proves the high application value of this research,but also verifies the correctness of the proposed algorithm theory from the perspective of practical application.