Research And Application Of Instance Segmentation With Enhanced Boundary Information

Instance segmentation,as the most challenging of the four classic tasks in computer vision,requires pixel level classification of each object in the image,because this object belong to different instances of the same category,Instance segmentation is very difficult.In reality,due to lighting,motion,and shooting angle,the boundaries of the segmentation mask are relatively blurry;At the same time,because of the complexity of the scene,the occlusion or the overlap between multiple objects lead to object disconnection and further increase the difficulty of the task.The current object detection and instance segmentation algorithms introduce a Feature Pyramid Network(FPN)structure for feature fusion to improve the detection and segmentation accuracy of small targets.They use the hierarchical features of the different resolutions to perceive objects of different sizes.By reason of information differences between low-level and high-level features,the feature fusion used for detection and segmentation tasks is not sufficient.There are few samples of damaged ancient inscriptions,which have the characteristics of single texture,diverse structure and semantics.Existing algorithms only detect and segment the entire character,but if the character is damaged,the detection effect is poor.the spatial structure analyzing of the character can provide effective support for the repair and recognition of damaged ancient inscriptions.The detection and segmentation of character components is an essential part of the spatial structure analysis of the character,and there is almost no related research abou them.Boundary information can provide powerful shape representations,which is crucial for enhancing object structure information,solving object boundary blurring and object occlusion problems.The purpose of research and application on instance segmentation with enhanced boundary information is to enhance the boundary and structural information of instances for improving the segmentation accuracy of masks and solving problems such as boundary ambiguity:At the same time,by integrating multi-scale hierarchical features,the problem of insufficient feature fusion is solved and feature representation is improved;Finally,in response to the problem of detecting damaged ancient inscriptions,this article proposes the detection and segmentation of character components,and enhances their structural information based on the characteristics of the inscription characters.Therefore,this article conducts research on boundary information enhancement and hierarchical feature fusion to improve the accuracy of the model’s boundary segmentation,and applies it to character component segmentation to explore its effectiveness in enhancing the boundary and structural information of character components,in order to achieve good segmentation performance of inscription components.The main research ideas are as follows:(1)Research on enhancing boundary information.Two traditional image processing techniques are introduced to enhance boundary information.The Sob el edge prior enhancement(SEPE)module uses the channel self-attention mechanism to capture the semantic dependency between cross channels by introducing the edge prior information extracted by the Sobel edge extraction operator to enhance the semantic information such as the boundary and structure of the pre-segmentation mask.In the boundary expansion refinement(BER)module,the dilation algorithm expands the boundary range of the pre-segmentation mask,and then extracts the features of boundary point and classify them again to correct the mask boundary wrong classification.(2)Research on hierarchical feature fusion.The main idea of the hierarchical fusion module(HFM)is to fully integrate the high-level semantic information and the underlying finegrained information by fusing all the hierarchical features of the previous stage after FPN.At the same time,because of this connection mode,the loss gradient of high-level features can be fed back to all levels of the previous stage when the loss is reverse transferred,thus improving the feature extraction ability of the underlying convolution of the backbone network for large objects.(3)Research on character component segmentation.Firstly,based on 336 character components,a character component dataset is generated by simulating the chaotic background of inscriptions using data augmentation;Secondly,the segmentation effect of existing instance segmentation models on character component datasets is explored,and it is found that a multiclass instance segmentation framework based on ROIAlign and ROIpooling operations can lead to component degradation;Finally,the Sobel edge prior enhancement(SEPE)module and hierarchical fusion module(HFM)proposed in(1)and(2)are applied to character component segmentation to explore their effectiveness in character component segmentation.(4)Finally,this article demonstrates the effectiveness of the proposed method through ablation experiments and comparative experiments with mainstream methods.The applicability of the proposed character component segmentation method has been demonstrated through visual representation,providing effective support for the restoration of damaged ancient inscriptions.