Material Surface Recognition Based On Joint Features Dimensionality Reduction Multi-Modal Extreme Learning Machine

In recent years,with the rapid economic development,the living standards of people have been continuously improved,and so as the demand of shopping.However,there are a wide variety of species materials on the market mixed with good and bad.People cannot accurately evaluate the quality of materials while selecting,so they may buy goods with quality problems.It is significance to explore a set of material surface recognition devices and algorithms for material surface recognition in response to this phenomenon.Visual can reflect the direction,thickness and interweave degree of the material surface texture,the material surface can be recognized based on the visual features.However,the visual cannot obtain finer features,such as the degree of roughness and the undulation and depth of the texture,which can be gained through tactile.Compared with single modality information,multi-modal information fusion can achieve complementation of each modality resource and more comprehensively recognize material surfaces.Therefore,starting from tactile or visual single modality,this paper designs material surface tactile and visual data acquisition devices,explores and summarizes a set of multi-modal recognition algorithms for material surface recognition,and create a fabric surface tactile information-visual image dataset for experimental verification.Firstly,this paper introduces the current research status of tactile,visual,and multimodal information fusion technologies in the field of material surface recognition,and analyzes the advantages and disadvantages of using tactile or visual information alone for material recognition,and draws the necessity,rationality,and innovation of applying the modal fusion theory proposed in this paper to material surface analysis.Secondly,this paper designs tactile and visual data acquisition devices,and explores and develops a set of algorithm processing flow for material surface tactile signals,including high-pass filter design,DFT321 three-axis tactile signal synthesis,and APSD power feature extraction and dimensionality reduction of the one-dimensional acceleration.After that,this paper introduces the texture feature extraction method of visual images.Thirdly,this paper explores a fusion method of tactile-visual joint feature dimensionality reduction for material surface recognition,uses the canonical correlation analysis?CCA?idea to randomly pair two modal samples,and uses dimensional reduction power features to express tactile time series and the texture statistics feature to represent the visual images,joint features to map to low-dimensional linear spaces,multi-modal extreme learning machine is trained to achieve classification.An 18-fabric dataset FTS-18 is created to verify the algorithm.The results show that the tactile-visual fusion information recognition effect is obviously better than the single modality.Further improve the algorithm and improve the random pairing pattern to a one-to-many pairing pattern,the classification advantage is more obvious.Finally,this paper verifies the validity of the multi-modal material classification algorithm on public LMT₁08 material texture dataset.