Neural Networks Learning For Large Scale Image Retrieval And Classification Problems And Its Applications

In this age of the internet, video and image seem to dominate the information media. Theimage traffic over the internet increases beyond any bound. For instance,2.5quintillion(2.5×10¹⁸) bytes of data were created daily in2012. This "Big Data" phenomenon is mostlycaused by image and video data. However, the lack of source of control point clearly creates aproblem for the quality and content of images being uploaded on the internet. Other problemssuch as transfer failure, intentional or accidental cover up of major content and noise will leadto missing of parts and information in images. Current neural network learning techniques forimage classification and retrieval problems which depend on training samples only, sufferfrom overfitting easily. Moreover, when the number of images increase rapidly, the selectionof training samples could become very inefficient.Therefore, in this work, we propose a sensitivity measure （SM） based framework forneural network learning to deal with aforementioned problems. Firstly, a LocalizedGeneralization Error Model （L-GEM） based on SM is proposed for a MLPNN to measure itsgeneralization error for unseen samples located near training samples. Then, we combine theSM with a feature deletion based learning algorithm for image classification problems withinformation loss in images or missing parts of images. On the other hand, gradients can notproperly propagate to deeper layers in a multilayer neural network with more than2hiddenlayers, so we proposed a new bi-firing activation function to relieve the gradient diffusionproblem in deep neural network learning. In order to meet the sublinear search requirementfor image retrieval in a large scale image database, a new hashing method is proposed basedon image filtering using the SM.The four major contributions of this thesis are as follows:1) A Stochastic SM （ST-SM） is proposed as a new penalty term for MLPNN training toachieve better generalization capability. The ST-SM measures the expectation of squaredoutput differences between training samples and unseen samples located within theirQ-neighborhoods for a given MLPNN. It provides a direct measurement on the MLPNN’soutput fluctuations, i.e. smoothness. Then, a2-phase Pareto-based multiobjective trainingalgorithm for minimizing both the training error and the ST-SM as objective functions is proposed to find the optimal architecture of MLPNNs.2) Error can not propagate efficiently in deep neural networks （more than2hidden layers）using current activation functions which have a large saturation regions in the input space.To address this problem, we propose a bi-firing activation function which is adifferentiable function with a very small saturation region.3) In applications like face recognition and handwritten digit recognition, feature valuemissing or missing parts of an object is common at the testing phase. Therefore, a newtraining algorithm （L-GEM-RFD） combing the L-GEM based on the ST-SM and a randomfeature deletion is proposed. The L-GEM-RFD minimizes the localized generalizationerror of neighborhoods of both training samples with full set of features and those withdeleted features. Experimental results show that the L-GEM_RFD yields better resultsthan state-of-the-art methods for image classification problem with images consisting partof objects being intentionally covered by a black block.4) Hashing is a sublinear method for image retrieval and classification. However, currenthashing method suffers from low precision and recall rates. So, in this work an imagefiltering method based on ST-SM is proposed to reduce irrelevant images from thereturned candidate images of mutli-hash method to enhance its precision while keeping itshigh recall.