Key Technologies And Applications In Cross-domain Image Recognition

Researches of supervised learning are usually carried out under the assumption that the training data and test data are from the independent and identical distribution(IID).Thanks to the construction of large-scale datasets in various application fields in recent years,neural networks with deep learning as the research theory has significantly promoted the development and application of artificial intelligence technology in computer vision,natural language processing,biological information computing and other fields.Among them,computer vision tasks represented by image recognition,such as image classification,object detection,semantic segmentation,etc.,have achieved better performance than traditional algorithms on many public datasets.However,in practical application scenarios,the distributions of real test data(target domain)and training data(source domain)are often different,which makes it challenging to generalize models trained on the source domain under the IID assumption to the target domain.Moreover,it is expensive and impractical to construct datasets with annotation information for target domains in cross-domain tasks.Therefore,there is an urgent need for research on how to solve the problem of image recognition with cross-domain distribution discrepancies in the field of deep learning.In this paper,we analyze the progress and limitations of existing research on domain adaption for cross-domain image recognition.Further,metric learning theory,adversarial learning theory and self-training learning are used as the theoretical bases,and feature condition distribution discrepancy,feature scale invariance,and feature structure correlation are taken as the research entry points,respectively.In several key crossdomain image recognition tasks(image classification,target detection,and semantic segmentation),a series of domain adaptive algorithms based on deep learning are intensively studied,and the related theories and techniques are applied to engineering applications.The main work of this paper is as follows:1)Domain category invariant feature learning based on conditional distribution alignment.Reducing the discrepancies in the conditional distribution of features will help domain category invariant feature learning,which can improve the generalization of the classification model to the target domain.In this paper,we proposed a novel domain adaptation algorithm called joint category-level and discriminative feature learning.Firstly,the learning of category feature representations is achieved based on source domain data with supervised information.Further,minimizing the discrepancies between sample features and category feature representations of the same category to achieve sample-level conditional distribution alignment;discriminative feature learning is achieved based on maximizing the discrepancy between sample features and category feature representations of different categories.In addition,the construction of a migratory weighting module based on domain classifiers is proposed to further encourage the model to obtain a better classification plane in the target domain.Experimental results show that the proposed method can significantly reduce the cross-domain distribution discrepancies and improve the classification performance on multiple cross-domain image classification tasks.2)Domain-scale invariant feature learning based on multiple adversarial learning.In more complex cross-domain target detection tasks,the differences in domain distribution due to the diverse scales of objects also can cause domain shifts.In this paper,a multi-scale feature enhanced domain adaptation model based on adversarial learning is proposed,which consists of two modules:a)Multi-scale feature fusion alignment module utilizing the contextual information conveyed among network layers to achieve scale-invariant feature learning;b)Multi-scale consistent regularization module to jointly optimize multi-level and multi-scale fusion feature learning,thereby encouraging the model to obtain more consistent domaininvariant features in the multi-adversarial learning.By constructing multiple cross-domain high-voltage transmission line inspection datasets,the proposed method significantly improves the performance of cross-domain object detectors on various tasks.3)Self-training based source-free domain adaptative method.In real-world application scenarios,it is a common phenomenon that source domain data is subject to privacy protection or computing resource constraints(such as autonomous driving,medical image analysis,etc.).When the source domain data is inaccessible,it leads to the inability to explore the domain-invariant feature representations of the source and target domains.Based on self-training theory,this paper is oriented to the cross-domain semantic segmentation task and exploits the source domain model prior knowledge to explore the target domain feature structure correlation representation.Specifically,we propose a pixel level consensus affinity representation to describe the structural correlation between paired pixels,and further form a consensus affinity loss to encourage the target model to generate consistent and relevant structured predictions between the nearest leading pixels.In addition,in order to suppress the unreasonable prediction of the source domain to the target domain,a pixel location-dependent uncertainty weighting scheme is proposed to encourage models that can learn consistent structured knowledge from reliable predictions of the target data.Experiments report the performance of the proposed method in four cross domain semantic segmentation tasks:synthesis to reality and reality to reality.4)Engineering practice of cross domain anti-counterfeiting drawing label identification.Facing the industrial research needs of anti-counterfeiting label identification based on computer vision,this paper introduces the overall research architecture of the project.Further,this paper analyzes the cross domain problems of key computer vision algorithms in the project,and introduces the two key technologies developed:a)"Blind super-resolution cross domain image generation model"|,which designs an adaptive method for blind super-resolution generation model guided by differences in visual feature distribution,solve the learning of blind super-resolution model under the condition of limited real low-definition images,so as to generate high-quality and high-definition images in cross domain tasks.b)"Structure perception guided cross domain similarity learning method",which potential estimation methods for structural similarity distributions in a single domain are investigated using feature structure correlation;and furthermore,structure-aware guided cross-domain contrast loss is formed to encourage distribution consistency learning between different domains for generalization enhancement of cross-domain image discrimination models.The research results are of great significance to promote the application of new anti-counterfeiting technology with deep learning as intelligent recognition algorithm in industrial scenes.