Research On Symmetric Target Detection Technology In Image Reflection Based On Salient Symmetry Feature

Symmetric objects are ubiquitous in natural images and synthetic images.As one of the most important global features of symmetric objects,symmetry detection has long been a research hotspot in the field of computer vision,and has been widely used in the tasks of image semantic extraction,image semantic understanding and emotion recognition.The detection technology of symmetric objects is to extract the symmetric target information contained in pictures.Generally speaking,it can specifically include extracting the position,symmetry category(such as translational symmetry,rotational symmetry,reflective symmetry,etc.)and the geometric information contained therein,such as the symmetry axis of reflective symmetry.The traditional symmetry detection method is feature-centered.As the name implies,this method is based on one or more extracted image features.Specifically,it first calculates the universal feature representation(such as SIFT feature)of the image,and then estimates the parameters needed by the symmetric target according to the characteristics of the feature.The advantage of feature-centered method is that it is convenient and quick,and the third party’s feature expression is enough to abstract the representation of the image.However,the disadvantage of this kind of method is obvious,that is,it needs to optimize a large number of calculated and selected feature matching pairs,which is also the main difficulty to be solved by this kind of method at present Moreover,it is worth noting that the importance and particularity of features are not taken into account in the construction of features due to the neglected role of feature engineering in symmetry detection tasks,which leads to the neglect of the semantic gap between lower-level and higher-level visual detection tasks.To solve the problem that traditional symmetric detection methods ignore the optimization of feature expression and matching pairs,this thesis proposes a complete framework from feature design to detection optimization.Specifically,this thesis improves the existing feature-centered method to make it more suitable for the detection of reflective symmetric targets and the extraction of related symmetric information.The improvement of this method mainly includes:1.Try to extract the feature representation under the specific task of reflection symmetry in a quantitative way,that is,Salient Symmetry Feature(SSF).The inspiration for designing this feature comes from human being’s sensitivity to symmetry.We try to extract a set of Salient symmetry points(SSF)in scale space to simply simulate the process of human perception of symmetric objects,and then encode the final SSF pixels with local information to obtain the final feature representation of symmetric objects.2.For images with unknown information,the size of symmetric kernel is very important.The size of symmetric kernel is small,and the loss of local information makes the measurement of symmetric response of salient points less,which leads to some important features being filtered out.When it is large,although more symmetrical salient point patterns can be detected in theory,it will also bring problems such as a sharp increase in computation.In order to solve these problems,we embed an Adaptive Kernel Size(AKS)algorithm in the framework to make the algorithm sensitive to symmetric kernel,so as to adaptively change the size of symmetric kernel.3.In order to eliminate noise error and match symmetry points more accurately,a matching symmetry point updating method based on random search and neighbor points is designed.At the same time,we will put forward that the symmetry axis information is explicitly transferred to a visual binary matrix-Symmetric Transformation Matrix(STM).With the help of STM rank matrix and the proposed energy function,the final symmetry axis of the image can be determined.The proposed symmetry detection framework is tested on two different datasets.The first experiment evaluates the effectiveness of the SSF proposed in this thesis.The experimental results show that when Loy’s method replaces the original features with the SSF in this thesis,the effect is slightly improved in the case of single axis,while it is relatively more obvious for multi-axis real images.Similarly,replacing the Nagar method feature with SSF improves performance even more in the single-axis case.In the second experiment,which is the evaluation of the overall symmetrical object detection effect,we compare the proposed optimized detection framework with the current state-of-the-art methods.The results show that the proposed framework can achieve the second highest performance in the case of single-axis and multi-axis,and there is no significant difference from the optimal method.For uniaxial reflection symmetry,this method is superior to other methods except Nagar.For multi-axis case,the proposed method achieves the leading performance,which also shows that the whole detection framework has a certain competitiveness compared with the existing detection methods.