| With the advancement of key technologies in various domains of the haptic internet,remote interaction with real-time objects and the perception of immersive operating experiences have become achievable.Through wireless communication platforms,users can manipulate and control both physical and virtual objects in various scenarios,utilizing haptic representation devices as displays to enhance the sense of immersion.However,the emergence of the haptic internet has also introduced new requirements in several areas,including the acquisition,compression,transmission,and display of haptic texture features.This thesis specifically focuses on one of these areas,namely,the acquisition of haptic texture features.Currently,the predominant methods for acquiring tactile texture features rely on physical devices,necessitating specialized acquisition equipment and skilled operators.This approach is labor-intensive,time-consuming,and incompatible with the real-time demands of haptic internet interactions.In recent years,using surface material texture images for haptic information extraction has gained attention.This method allows for direct extraction of haptic texture features from images,providing a faster and more convenient alternative to collecting tactile texture features.However,existing research on haptic data acquisition from images has limited progress in meeting the requirements of the haptic internet.Some studies focus on generating basic haptic data like friction coefficient,while others generate unsuitable acceleration signals for reproducing surface haptic textures.Privacy concerns and user reluctance to exchange haptic data within the haptic internet can lead to underfitting issues when training haptic texture feature acquisition models.This thesis addresses these challenges by acquiring tactile texture features from visual texture features of surface materials.It employs machine learning techniques like stacked auto-encoding theory,gated recurrent networks,and federated learning to construct a generative model for tactile texture features.(1)The feature extraction method of visual texture images of surface materials is researched based on three advanced image processing network structures,e.g.,deep convolutional neural network,residual network and Mobile Net.The validity of the extracted visual texture image features of surface materials is also verified by simulation experiments.(2)This thesis proposes a tactile texture feature acquisition method for surface materials based on stacked auto-encoder theory and gated recurrent networks.Specifically,the tactile texture features of surface materials are extracted by stacked auto-encoder,and then a state transition model from visual texture features of surface materials to tactile texture features is established in low-dimensional space by gated recurrent networks.The parameter-stabilised model takes as input a texture image of the surface material to obtain tactile texture features,avoiding the problems of complex acquisition equipment and high time costs involved in traditional methods based on actual device measurements.(3)Combining federal learning with tactile texture feature acquisition for surface materials,the proposed distributed tactile texture feature acquisition method for surface materials is able to solve the performance degradation problem in the tactile Internet due to insufficient user data.It also does not require the user to transmit the original data,thus protecting the user’s privacy and data security.This thesis investigates the method of acquiring tactile texture signals from surfaces in terms of the underlying theory,model setup,method selection and simulation validation to provide a reference for the practical application of acquiring tactile texture information from image texture information of surface materials in the haptic Internet. |
