Mobile Visual Search Research And Optimization Based On Extreme Learning Machine

With the popularity of mobile devices, mobile terminal has become an image andvideo processing equipment with many very powerful features, such as highresolution, color display, etc. Mobile visual search refers to combining the actualimage collected by mobile terminal with the internet search technology to detectobject and finally display the related resources about the retrieval image. Thoughthere has been many researches and explorations on this technology, it has not beenwidely used due to environmental conditions. In this paper, the ELM (Extremelearning machine) is applied to the mobile visual search direction and some relatedresearch has been done to analysis the feasibility of mobile visual search.Mobile visual search is based on image recognition technology. Traditional imageretrieval techniques usually use the SVM (Support Vector Machine), neural networkmethod to train the image characteristics data, which need multiple iterations andprone to produce local optimal solution. As a new technology, Extreme learningmachine (ELM) overcomes some problems that the other traditional intelligencetechnology faces. It is a simple learning algorithm, different from the traditionallearning algorithm, such as the neural network learning algorithm. ELM training cannot only achieve the smallest error, but also be committed to minimum standardweight at the same time. It avoids the possibility of slow learning and failure ofnetwork training which traditional method SLFNs face.In terms of theory, this paper first combined ELM algorithm and the derivedmathematical formula, verified and analyzed the ELM’s ability of dealing withclassification, then extended the ELM to the image area. On this basis, this paper putsforward the BMVS(Basic Mobile Visual Search) frame model, applies the ELM onserver side of this frame and researches the Mobile visual search by image acquiringof Mobile devices and ELM processing on network side.Then in the process of validation and analysis on the server-side, this paper putsforward A-ELM (Ascending-Extreme Learning Machine) algorithm to reduce therandomness when ELM is doing classifying. This algorithm avoids random settings tohidden layer neuron which is happened in the process of ELM’s classification, whichmakes the training and testing in the classification process more balanced. In the experiment, this paper focuses on the verification of the feasibility ofBMVS model, including the experimental treatment on the server side andexperimental display on the mobile terminal. Finally, this paper summarizes theexperiments to prove the reliability of ELM in mobile visual search and concludes theefficiency of the ELM by comparing with the SVM’s experiment.