Research On The Algorithms Of Image Classification And Recognition Based On Feature Learning And Machine Learning

With the development of the multimedia Internet era,the network data is growing at an amazing speed.In the face of massive image data,how to use computer image classification and recognition technology to quickly distinguish and effectively manage them is becoming more and more important.As one of the important means of computer image processing,computer image classification,and recognition technology has been a hot topic in recent years because it spans the fields of image processing,computer vision,machine learning,and pattern recognition.Because the collected image data will be affected by factors such as illumination,Angle,and diversity of similar targets,image classification,and recognition technology is still a challenging research problem.The key to image classification and recognition is how to get efficient image representation.While traditional image feature learning frameworks have researched for a long time at the early stage,researchers have switched their interest to Deep Learning based image classification with the development of Neural Network.Image classification in research and application filed is effected by the image difference between inter-class and intra-class.This article will focus on the problems in this technique,using feature and machine learning methods,from three scientific points of visual feature,extreme learning machine,and deep learning academic research perspective.We also investigate three practical applications of supervised learning based on single-label and multi-label,semi-supervised learning based on noise label and unsupervised learning based on pedestrian recognition.The main contributions of this paper as follows:(1)Firstly,we focus on the feature description,feature encoding,and machine learning for the traditional image classification problem.We present an image classification algorithm based on visual feature learning and extreme machine learning method.Most traditional image classification systems depend on the gray-based SIFT descriptors,which only calculate the gray layer variations of the images.However,the ignorance of the color information may lead to misclassification of images.In this paper,an image classification method based on different kinds of colored SIFT descriptors is introduced and implemented.Then,this paper uses the local constrained linear coding and local soft assignment coding as feature coding methods to further improve the performance of the image classification framework.Finally,this paper brings in a new ELM learning algorithm to compare our framework to the traditional classification framework.In order to achieve better performance,we also present a new image classification system by integrating color SIFT descriptors and feature coding method with the extreme learning machine method.(2)Secondly,we focus on the convolutional neural networks(CNNs)in image classification.We propose multiple vectors of locally aggregated descriptors(VLAD)encoding method with the CNNs features for image classification.Despite the effectiveness of CNNs especially in image classification tasks,the effect of convolution features on learned representations is still limited.It mostly focuses on the salient object of the images but ignores the variation information on clutter and local.We propose an image classification framework,which is the VLAD encoding method with the CNNs features.Furthermore,in order to improve the performance of the VLAD coding method,we explore the multiplicity of VLAD encoding with the extension of three kinds of encoding algorithms.We equip the spatial pyramid patch(SPM)on VLAD encoding to add the spatial information of CNNs feature,which leads to further improve the performance of image classification.(3)Thirdly,we focus on learning large-scale noisy labeled data for image classification and recognition.We propose an algorithm of product image recognition with guided learning and noise supervision.Following the development of an intelligent retail online shopping market,this paper considers new research on real and daily product image recognition.This is not only an important problem in real-life applications but also a very challenging practical problem.In order to explore and solve the task of product image recognition in the real world,we introduce a novel large-scale product image dataset,called product-90.The image data and noisy label data are downloaded from the reviews from searched products by given key names on e-commerce websites.The images in the consumer review are casually captured and labeled.Therefore,these image data are often affected by background clutters,category diversities,and noisy labels.This paper presents a general guided learning method for mass noisy labeled data.To advance the performance of daily product image recognition,we propose a generic guidance learning method to take full advantage of the small clean subset and the largescale noisy data in Product-90.(4)Fourthly,we study the label dependencies for the multi-label image classification and recognition.We proposed a unified multi-label image recognition framework based on the adaptive graph convolution network(A-GCN).The A-GCN leverages the popular Graph Convolutional Networks with an Adaptive label correlation graph to model label dependencies and train the label-dependent object classifiers.Because some objects usually appear together in the real world,exploring how to effectively capture the dependencies between object labels is the key to a multi-label image recognition problem.We introduce a plug-and-play adaptive Label Graph(LG)module to learn label correlations with word embeddings,and then utilize traditional GCN to map this graph into label-dependent object classifiers,and further applied these classifiers to image features.In addition,we propose a sparse correlation constraint to enhance the LG module.Extensive experiments show that our A-GCN achieves superior performance on multi-label recognition.(5)Finally,we study the unsupervised person reidentification(Re-ID)task.It still faces the challenge to learn robust and distinguishing feature representations.To alleviate this problem,we propose a conceptually simple yet effective block for unsupervised person Re-ID,termed as Meta Feature Attention(MFA).MFA mainly learns sample feature interactions for small groups with an attention mechanism in each mini-batch and generates a new sample feature for each group by weighted sum.We train the whole network in an end-to-end fashion on two independent domains by multi-task learning.The main benefits of MFA come from two aspects:1)numerous new samples can be used for training which densely extends the feature space and makes networks better in generalization;2)the trainable attention weights capture the importance of samples which make networks focus on more useful or discriminative samples.