Object Recognition Under Complex Environmental Conditions on Remote Sensing Image

Transcription

1 International Journal of Applied Environmental Sciences ISSN Volume 11, Number 3 (016), pp Research India Publications Object Recognition Under Complex Environmental Conditions on Remote Sensing Image Jian-min Liu* 1, and Min-hua Yang 1 1 School of Geosciences and Info-Physics, Central South University, Changsha, 41000, China. School of Information Science and Engineering, Hunan Institute of Humanities Science and echnology, Loudi , China. Abstract Object detection on high-resolution and optical remote sensing images has attracted extensive interest. Because Deep neural networks have boomed in recent years, based on caffe and cuda, We trained a multilayer convolutional neural network to classify the 1 thousand high-resolution and unlabeled optical remote sensing images via the Internet into the 1000 different classes. On the test dataset, we achieved error rates of 17.7% which work went relatively well than the previous machine learning techniques. o help to save training time, we also used pre-trained model, GX970 GPU and powerful implementation of the algorithm, these approach helping we saving a lot of time. he results compared with the published ones, and good agreement is obtained. Keywords: Deep Neural Network(DNN); Remote Sensing Image; Detection 1. Introduction In military techniques and many other fields, aircraft detection is increasingly used. Various types classification methods have been applied to remote sensing work as object classifiers. Conventional remote sensing image detection techniques mainly based on spectral, and the detection algorithms typically include linear regression, nonlinear regression, logistic regression, k-means, conditional random fields and maximumlikelihood[1][]. hose like above can be divided into the method is based on statistical methods and non statistical methods. Firstly, statistical based method y x; ~ (, ) can be divided into regression problem and classification problem y x; ~ Bernoulli( ).μandφthat is said that x 和, both of them look different, but in

2 75 Jian-min Liu and Min-hua Yang fact, are two special cases of general linear model with a model that Michael.I.Jordan proposed. Mathematical expression can be written as: p( y; ) b( y)exp( ( y) a ( )) (1) Extending from logistic regression method, it is very common that classification among variety objects except two classification method in reality, such as remote sensing image target for farmland, soil, water, road...,let y {1,,..., k}, then p( y; ) exp(((y)) 1 log( 1 / k) ( ( y)) log( / k) ()... ( ( y)) log( / ) l og( ) let 1 k1 k1 k k log( / k ) log( / ) k, a( ) log( k ),b(y) 1 log( k 1 / k) hen (1) is expressed as: p( y; ) b( y)exp( ( y) a( )), and it is called softmax algorithm. On the one hand softmax algorithm can be understood as a generalization of the logistic algorithm, but on the other hand the logistic algorithm is a special case of the softmax algorithm. he above methods is based on the statistical analysis of the spectral and spatial information of each pixel. Secondly, non statistical methods. based method typically include Neural network and Support vector machine algorithm, etc. In the face and handwritten numeral recognition, remote sensing image classification and other applications, it is one of the best machine learning methods before deep learning boomed. he researchers made a lot of results by using spectral information, spatial information and object information. On many occasions, SVM-based classifiers work went relatively well than than other extensive used machine learning techniques[][3]. On the MNIS dataset handwritten digit recognition task, the highest-accuracy reports show backpropagation hit 1.6% of error rates only if the initial weights has been initialized randomly and support vector machines hit 1.4% of error rates[3]. Owing to the complexity a long time, these classifiers has been locked in the state-of-theart methods[4][5]. Given these issues, we consider the challenge of character and digital recognition, aircraft detection and a gradually improving understanding of the human's feelings[6]. In years gone by the expectant, based on hand-engineering approach these issues are certainly laborintensive, and work inefficient for new problems. Some nonsupervised feature representations ones was expected to replace hand-engineered methods[4][6]. DNN algorithm like sparse autoencoder, deep boltzmann machines and deep convolutional neural network method equip scientist and R&D engineers to classify, detect and recognize various types objects. A large, deep neural network, such as sparse autoencoder nonsupervised learn features from high-dimensional original data, and train central layer which consist of a lot of hidden units to reconstruct original input vectors, and detect edges at key positions and structure of the image[4][6][7]. o judge the weights of each unit of each layer, gradient descent is susceptible to local fine-tuning in such a autoencoder network. Speaking from experience gradient descent works well only if the initial weights has been initialized outline low-level structures such as edges from it, based on this, and form a just and comprehensible view of high-level structures such as corners, structure, and outline. he feature acquired from unsupervised learning can be used to reconstruct original input (3)

3 Object Recognition Under Complex Environmental Conditions 753 vectors[8]. After fine-tuning, a network with three hidden layers forms a very good generative model of the joint distribution of handwritten digit images and their labels. his generative model gives better digit classification than the best discriminative learning algorithms[9]. Speech recognition is a classic application of neural networks, in recent years, and deep learning has achieved remarkable results in speech recognition in recent years. Microsoft and Google have been gradually put deep learning's speech recognition algorithm into commercial applications. Microsoft's original speech recognition based on the use of hybrid algorithm Gauss reduce the error rate of words to 16%. Bengio, Hinton and other scholars solved MNIS based classification problems, and breaking the old mark 0.14% error rate Set by traditional machine learning algorithm for many years[10][11]. Krizhevsky had made significant progress on the ImageNet data set with a 15.3% error rate[1]. he performance of traditional machine learning methods rely heavily on the choice skills representation or feature method of object data, which can be called an art of scientific research rather than a technology, need to be trained with regularity by researchers and is extremely strict by application environment. So the actual effect of traditional machine learning algorithms depend on pretreatment technology and data conversion which whether can support the machine learning algorithm. he process of selecting such characteristics is important but time-consuming, and algorithms can extract and organize data from independent primary characteristics. he implementation of production process characteristics are selected from massive data based on human intelligence of researcher's decision instead of training with regularity unsupervised operation. Human intelligence can make up for the weakness of the algorithm. In order to expand the application field of the machine learning algorithm, reduce the labor cost, we must reduce the algorithm of feature selection dependence on trained researcher, and realize true artificial intelligence, automation and intelligent independent from the brain[13]. his paper presents aircraft detection on high resolution optical remote sensing images based on Deep Neural Network which consists of five convolutional layers, some of which are followed by max-pooling layers, and three fully-connected layers with a final 1000-way softmax. he results are compared with the published ones. Finaly, because GPU's memory available is scant, the dimension of network is limited to a certain location. Our network spends about 19 hours to train on one GX 750i GB GPUs.. HEOREICAL ANALYSIS A common Deep neural network consists of autoencoder layer, convolutional layer, pooling layer, and fully-connected layer. Similar to the conventional method, and for the sake of convenience, the equations of overall cost of backpropagation is: n mu - 1 sl s u+1 ( i ) ( i ) ( u) V, d i j j =1 u=1 j =1 i =1 1 1 γ C( V, d) =( k ( x - y ) ) ) + ( V ) n 4 compute partial derivation, and expressed as (4)

4 754 Jian-min Liu and Min-hua Yang V n 1 γ CV (, d) ( CV (, d; x, y ) ) V n C ( u) ( u) j i j 1 j i We obtain: CV (, d; x, y ) ( u) V ji CV (, d; x, y ) ( u) d ( j ) ( j ) ( u) ji a ( u) ( u1) i j ( u 1) j i (7) After that: C sp 1 (, ) s C V d log (1 )log ^ ^ i1 1 1 (( ( )) ( ) () () (3) ' s j Vij 0 g z ^ ^ i1 i 1 i i i (9) After that, the convolutional layer execute an image convolution for output of the previous layer, where the filter of convolution layer named learned core which consists 16x16 matrix of feature. he layer after fully-connected layer works with a final multi-way softmax. On this layer, we choose softmax model which generalizes logistic model in multi-class classification and recognition. ( j) ( j) p( y 1 x ; w) ( j) ( j) p( y x ; w) ( j) ( j) p( y 3 x ; w) ( j) kw( x )... ( j) ( j) p( y h x ; w) (10) Because summary of probability is equal to 1, we obtain w1 e w e w3 e ( j) 1 k ( ). w x h wi e. i1. wh e (11) We obtain a working model of softmax which come into play in multi-class classification and recognition. (5) (6) (8)

5 Object Recognition Under Complex Environmental Conditions Experimental results and analysis: In this Experiment, We focus on aircraft detection on high-resolution and optical remote sensing images. Based on caffe and cuda, we trained a multilayer convolutional neural network to classify the 1 thousand high-resolution and unlabeled optical remote sensing images via the Internet into the 1000 different classes. On the test dataset, we achieved error rates of 17.7% which work went relatively well than the previous machine learning techniques.. o help to save training time, we also used pre-trained model, GX970 GPU and powerful implementation of the algorithm, these approach helping we saving a lot of time. he process from the first layer to the seventh layer of the classification of an image is shown in Figure 1-(a-l). Figure 1-(a) show the input raw image, Figure 1-(b) show the output of first layer named conv1 layer, Figure 1-(c) show the filters of second layer named conv layer, Figure 1-(d) show the output of second layer named conv layer, Figure 1-(e) show the output of third layer output named conv3 layer, Figure 1- (f) show the output of fourth layer named conv4 layer, Figure 1-(g) show the output of fifth layer named conv5 layer, Figure 1-(h) show the fifth layer after pooling named pool5 layer, Figure 1-(i) show the first fully connected layer named fc6 layer, Figure 1-(j) show the second fully connected layer named fc7 layer, Figure 1-(k) show the output of final probability, and Figure 1-(l) show the top 5 detected labels. he top 1 detected label is space shuttle, the top detected label is aircraft carrier, the top 3 detected label is liner, ocean liner, the top 4 detected label: speedboat, and the top 5 detected label is yawl.

6 756 Jian-min Liu and Min-hua Yang Figure 1: Experimental results. (a) he input image (b)he first layer output, conv1 (c) he second layer filters, conv (d) he second layer output, conv (e) he third layer output, conv3 (f) he fourth layer output, conv4 (g) he fifth layer output, conv5 (h) he fifth layer after pooling, pool5 (i) he first fully connected layer, fc6 (j) he second fully connected layer, fc7 (k) he final probability output (l) he top 5 predicted labels. 4. Conclusion his paper on the one hand provide an extension to the existing analytical methods of aircraft detection on high-resolution and optical remote sensing images, on the other hand, it maybe help to further the development of this Deep Neural Network means. Detail See able 1. able1: Comparision of Convolutional Neural Network and traditional method Convolutional Neural Network raditional Method Complexity A highly complex multilayer Overall structure is not convolutional neural network is complicated, but Experienced constructed manual intervention is researchers and a lot of time are unnecessary needed Input Requirment Original unlabeled image Pre processed image Memory Sharply increase the memory Less memory requirement, but requirement requirement of training time more manual intervention. Does not deduce the memory requirement of working time Simulation Reduce CPU and GPU time Long CPU time is required time siginificantly. GPU not Available About from 1/3 to 1/10 time of that of traditional method. Acknowledgments Supported by Provincial Natural Science Foundation of Hunan ( 01JJ509 ) Supported by Science and echnology Planning Project of Loudi (014KJ04)

7 Object Recognition Under Complex Environmental Conditions 757 References [1] G. M. Foody and A. Mathur, A relative evaluation of multiclass image classification by support vector machines[j], IEEE rans. Geosci. Remote Sens., vol. 4, no. 6, pp , Jun [] Yushi Chen. Deep Learning-Based Classification of Hyperspectral Data[J], IEEE JOURNAL OF SELECED OPICS IN APPLIED EARH OBSERVAIONS AND REMOE SENSING, VOL. 7, NO. 6, JUNE [3] J. A. Gualtieri and S. Chettri. Support vector machines for classification of hyperspectral data[p], Proc. IEEE Geosci. Remote Sens. Symp. (IGARSS),Honolulu, HI, USA, 000, pp [4] Honglak Lee Chaitanya Ekanadham Andrew Y. Ng. Sparse deep belief net model for visual area V[P].NIPS 007. [5] L. Zhuo et al., Agenetic algorithm based wrapper feature selection method for classification of hyperspectral images using support vector machine[p], Proc. Geoinformat. Joint Conf. GIS Built Environ. Classif. Remote Sens. Images Int. Soc. Opt. Photonics, Nov. 008, pp J 71471J.]. [6] Honglak Lee, Roger Grosse, Rajesh Ranganath and Andrew Y. Ng. Convolutional deep belief networks for scalable unsupervised learning of hierarchical representations, ICML 009. [7] G. E. Hinton* and R. R. Salakhutdinov Reducing the Dimensionality of Data with Neural Networks SCIENCE VOL JULY 006 [8] Honglak Lee, Roger Grosse, Rajesh Ranganath, and Andrew Y. Ng. Unsupervised Learning of Hierarchical Representations with Convolutional Deep Belief Networks.Communications of HE ACM VOL.54 NO.10, [9] Geoffrey E. Hinton A Fast Learning Algorithm for Deep Belief Nets.Neural Computation 18, (006)] [10] G. E. Hinton,S. Osindero,Y. W. eh.a fast learning algorithm for deep belief nets. Neural Computation. 006 [11] Y. Bengio, G. Mesnil, Y. Dauphin, and S. Rifai, Better Mixing via Deep Representations, Proc. Int l Conf. Machine Learning, 013. [1] Krizhevsky, A., Sutskever, I. and Hinton, G. E. ImageNet Classification with Deep Convolutional Neural Networks. NIPS 01: Neural Information Processing Systems, Lake ahoe, Nevada [13] Liu Jianmin, Huang Fan, Dai Jun. Image recognition technology for remote sensing based on research and application Machine learning. Journal of Xi'an University of Arts & Science. Oct. 015

8 758 Jian-min Liu and Min-hua Yang Authors LIU Jian-min, PhD student Research interests: Remote Sensing, GIS, and Machine learning,