Research Of Improved Multi-layer Extreme Learning Machine And Application Of Automatice Valuati On Of Selfies

Taking photos at any places and any time has become the daily habits of most people due to the intelligentization and popularization of mobile devices,especially,with the rapid development of social networks,sharing selfies become a fashionable way to show yourself and express inner.Taking selfie photos has become a global hot phenomenon.Because of the different perspective of aesthetics,it is difficult to judge whether a selfie photo is popular or not.So it is meaningful to study and propose a algorithm for automatic evaluation of selfie photos before people post their selfies.It may reduce the negative impact of posting selfies blindly.However,automatic evaluation of selfie photos is challenging problem due to the different perspective of observers,background,illumination,expression,camera angle,occlusion.Using data driving method and deep network model learns the public aesthetic.The establishment of an effective evaluation algorithm has theoretical and protical significance.It not only provides the solution for the automatic evaluation,but also provides the research experience for the analysis of unconstrained expression.In order to be fair and objective to score selfie photos,this paper use a way of scoring selfie photos by many people.Based on the human eye paying more attention to the local important region of one image,we propose the algorithm of multi-layer extreme learning machine with enhancing the local significant region(ML-ELMELSR).The algorithm enhances the recognition contribution of local significant region.The local significant region is different for different tasks.The algorithm enhances the contribution of important local region in recognition.We usually use the central region or the regions of being detected by other algorithms as local significant region.Besides,we study the detection algorithm for the region of interest based on selective search,we turn detection task into binary classification task.And Experiments shows ML-ELMELSR uses less training time and achieves a better recognition result than other algorithms.The main research work in this paper is as follows:1、Building selfie photo dataset and establish image grading calibration system.There is no public open source image dataset in the research field of selfie photos.So it is necessary to build selfie photo dataset.The selfie photo dataset is the basis of our research.We get a large number of selfie photos by selecting the appropriate web crawler framework.And then we can obtain high quality candidate samples after data cleansing,duplicate removal,samples selection,normalization and other preprocessing operations.In addition,we need to score each photo of the dataset.we have finished the work of establishment of a calibration system and selfie dataset.2、Research on ROI detection algorithm based on selective search.The method of the region of interest detection is comparing the candidate region and the target region by using sliding window and the exhaustive search strategy.This method has significant shortcomings such as excessive computation,long time consuming,poor application and so on.In this paper,the method based on selective search turn the region detection problem into two classification problem.The initial regions are generated by the image segmentation,then merging the regions by calculating the similarity of two regions.So the candidate regions are generated.Finally,the positive and negative samples are obtained by calibrating the region of interest and the candidate region set.A classifier model is trained to determine whether a region is a region of interest or not.Then we can test the candidate regions generated by test images.How to define the candidate regions and the strategy of merging the regions is one of the emphases for research.3 、 Proposing the algorithm of multi-layer extreme learning machine with enhancing the local significant region.The extreme learning machine has the advantages of short training time and high recognition accuracy.The multi-layer extreme learning machine can extract the abstract feature of the image,and the ability of classification is more powerful.However,the algorithm is lack of knowledge of local significant regions.In this paper,we extract the local significant regions,use extreme learning machine auto encoder with the local significant regions to train the every layer of network.And stacking the extreme learning machine auto encoder with the local significant regions layer by layer become a deep network.Extreme learning machine auto encoder with local regions can provide the network more information,the generalization ability of the network can be more powerful.How to find and detect the local significant regions is another emphases for research.4、Application of multi-layer extreme learning machine on automatice valuation of selfies.We use data-driven approach to build a selfie dataset which scoring by mnay people.So the dataset will meet the public aesthetics.Comparing to several traditional algorithms.Using ML-ELM and ML-ELM-ELSR algorithms to learn the better representation of public aesthetics in the selfie dataset.And it will find the common in the good selfies and provide people reference before posting the selfies.