Research On Image Data Processing Methods Based On Deep Learning

In the contemporary era,the explosive growth of image data has posed higher demands on image data processing technologies.Image semantic understanding,the ability of computers to comprehend the content within images akin to human perception,stands as one of the pivotal tasks in image data processing.This dissertation delves into an extensive exploration and research of deep learning-based image semantic understanding methods,addressing commonalities and solutions for classification and segmentation tasks across various image modalities,including two-dimensional(2D),three-dimensional(3D),and point cloud images.The research zeroes in on deep learning,image processing,and semantic understanding methodologies,employing convolutional neural networks(CNNs),graph neural networks(GNNs),and attention mechanisms to tackle key issues such as multi-scale variations,subtle inter-class differences,data sparsity,and object-driven perspectives.The contributions of this dissertation are summarized as follows:(1)When processing 2D images,multi-scale variations and subtle inter-class differences are pervasive issues.Due to the varying sizes and distances of objects in the real world,objects of the same category may manifest in different dimensions within images,and the differences between different categories of objects are subtle.To address these challenges,this dissertation introduces an adaptive two-dimensional image semantic segmentation model that accommodates targets of varying sizes.Utilizing satellite image data as a backdrop,the model’s efficacy in handling building segmentation tasks within satellite imagery is validated through the construction of a rooftop segmentation dataset.(2)In the realm of 3D imaging,data sparsity emerges as a salient issue.Particularly in medical imaging,3D images often suffer from insufficient resolution,low contrast,and limited coverage due to limitations in imaging technology and equipment.In response to this challenge,the paper constructs an improved 3D image semantic segmentation model based on image projection networks.By effectively addressing issues of information concentration and data layering,the model enhances the performance of 3D image segmentation.When applied to medical image analysis tasks such as optical coherence tomography(OCT)of the eye,the model demonstrates significant advantages.(3)As an emerging image modality,point cloud images present new challenges to image semantic understanding due to their sparsity and object-driven perspective characteristics.The paper proposes a chain-like structure to effectively integrate global and local features,enhancing the model’s expressive power and flexibility.Through this flexible structure,the model performs efficient feature extraction across multi-scale dimensions.Additionally,by employing attention mechanisms to focus on key areas,the model improves the precision of identifying objects with subtle inter-class differences.In summary,this dissertation conducts research across multiple specialized domains within the field of image semantic understanding and has been validated and applied through experimentation.By resolving key issues such as multi-scale variations,subtle inter-class differences,data sparsity,and object-driven perspectives,this dissertation offers new insights and methodologies for research in image data processing and image semantic understanding.