Research On Methods Of Images Deep Feature Extraction

Feature extraction of images refers to learning feature representation from images,which has proven beneficial for image vision tasks.The representation ability affects the performance of image vision tasks directly.In recent years,deep learning has significantly promoted the generalization ability of feature representation and achieves better performance than traditional learning methods.It has been successfully applied to visual tasks such as image classification and object detection,even surpassing human behaviors in some fields.At present,the regime of deep learning focuses on two major aspects: network architecture and loss function.The former explores the number of convolution channels and layers,the combination of different convolution kernels,and the adaptive selection of different branches in convolution neural networks(CNNs).Generally,the network is built on the basic building block.In contrast,the latter is to measure the distances between the predictions of CNNs and the corresponding ground truth,serving as training deep models.Both aspects affect the performance of CNNs,to some degree.In fact,the academics make initial attempts from two aspects to leverage layer-wise semantic information,in which they underlie important theories and have wide application prospects.The current focuses of network architecture and loss function for image classification are attentional mechanism and cross-entropy loss,respectively.To this end,the main contents of this thesis are:1)a basic block of the new attention model is devised.Inspired by attentional mechanism of human visual system,attention strategy used in deep learning is to highlight the feature regions of interest.Most attention models belong to soft attention,which can be coupled with in deep networks in an end-to-end manner.It can be grouped into two categories: channel attention and spatial attention,which correspond to the reweighted operations over channels and features,respectively.In this thesis,the proposed attention model exploits each pixel of feature maps by itself to induces two attentional values through a nonparametric activation function and a simple convolution operation.This model can be seamlessly embedded into current CNNs,with the purpose of enhancing classification performance of two mainstream deep networks.2)a new center loss is investigated to avoid two limited aspects of the cross-entropy loss: singly utilize the intra-class distance but ignore semantical ambiguity about the inter-class distance;the generalization ability is weak.Specifically,the proposed center loss reduces the gap between samples of the same class whilst enlarging the distances between samples of different classes by respectively centralizing the deep features and incorporating the between-class scatter.Moreover,the proposed loss regularizes the probabilistic output of CNNs to boost the generalization ability of the network.Two mainstream deep convolutional networks are trained by our loss for image classification,and the performance is superior to existing loss functions.