Quality-Based Object Recognition

Daily registration through face recognition,pedestrian tracking and a series of applications have gradually rose in people's life,which shows that face recognition and pedestrian re-identification are playing more important role in practical applications.In recent years,although the existing algorithms have achieved high performance,these performance are mostly based on better quality images.However,in real life,the obtained images through various photographing devices often have serious changes on gestures,facial expressions and appearance as well as other problems such as significant light changes and occlusions.Quality varies among different images,poor quality images will have a negative impact on target recognition tasks.For video object recognition,quality evaluation can provide corresponding weights for features from each frame according to the image quality of each frame during feature fusion,so that the performance of the target recognition task can be improved.For image quality evaluation,in order to improve the accuracy of target recognition,the image quality evaluation results should be consistent with the target recognition results.Based on the combination of image quality evaluation algorithm and target recognition task,this paper mainly focuses on face recognition-oriented image quality evaluation and quality-based video target recognition algorithm.There are still two problems in the existing face quality evaluation algorithms.First,the existing quality evaluation algorithm uses the subjective quality evaluation(MOS value)as ground truth.Although the MOS value reflects the requirement of human visual system,it is inconsistent with the face recognition system.In view of this difference,some scholars use the similarity between the test image and the reference image to define the evaluation ground truth,but this method brings new problems: the unreliable artificial choice of reference image and limited practical application.Second,the features used by the existing algorithms are not robust.Existing algorithms based on neural network structure use only semantic features to evaluate the image quality,ignoring the local features.In order to solve the above problems,this dissertation has done the following work in the task of face image quality evaluation.Firstly,the neural network-based face recognition system is used to generate the ground truth of the image quality evaluation model to ensure the consistence between quality evaluation results and face recognition system results.Second,when evaluating the face image quality,this paper incorporates different scale facial features to enhance the robustness of the features.In addition,in order to reduce the time and space complexity of the regression model,the paper applies sparse representation to deep features.The experimental results show that the proposed algorithm can effectively reflect the recognizable degree of the image through quality score,which means that the algorithm can select the best face image and improve the performance of the face recognition system.Further,this paper studies the video object recognition algorithm based on quality evaluation.At present,the existing algorithms use the quality evaluation score as the weight of each frame feature,and finally,the weighted sum of each frame feature is used as the representation of target video feature to obtain performance improvement.However,there are two problems with these algorithms.Firstly,the feature extraction and quality evaluation algorithms used in the target recognition process extract features from each frame independently,ignoring the relevance between the frames.Secondly,these algorithms lack of robust characterization for the target features,failing to take into account both global and local features and resulting in incomplete characterization.In order to solve the above problems,we completed the following work: Firstly,we added recurrent network(LSTM)module to the feature extraction and quality evaluation module to mine the related information between frames and obtain more effective features.This paper considers the temporal information in videos and makes features more reasonable.Secondly,this article combines global and local features to obtain a more complete representation of the target.That is,it contains both the global structure features and partial bodily features.For face recognition tasks,we conduct experiments on the YouTube Face Database,using the Receiver Operating Characteristic curve to evaluate the performance of our face recognition algorithm.Experimental results show that our performance is higher than the existing algorithms'.When the false positive rate is 0.01,the true positive rate of our method is 0.78,while the highest true positive rate of existing methods is 0.549,which means that our method greatly improves the verification accuracy for positive samples under high threshold.For pedestrian recognition,we experiment on the iLID_VIDS and PRID2011 datasets.Experiments show that our top1 matching rate improves about 3% compared comparing with existing methods.