10
Intelligent Enterprise‐Level Big Data Analytics for Modeling and Management in Smart Internet of Roads

Amin Fadaeddini1, Babak Majidi1,2, and Mohammad Eshghi3

1Department of Computer Engineering, Faculty of Engineering, Khatam University, Tehran, Iran

2Emergency and Rapid Response Simulation (ADERSIM) Artificial Intelligence Group, Faculty of Liberal Arts and Professional Studies, York University, Toronto, Canada

3Computer Engineering Department, Shahid Beheshti University, Tehran, Iran

10.1 Introduction

The amount of machine‐generated data produced by a metropolis or a regiopolis is significantly larger than a city and manual interpretation of these data is not possible. In the past decade, machine learning algorithms provided the capability of modeling the big data streams and to produce patterns which can assist the decision‐makers. A metropolis is a major city combined with the smaller satellite cities around it. The increase in the population and gradual expansion of both the major central city and the smaller cities around it led to a combined urban environment which is referred to as the metropolis. As the living costs make the central urban areas unaffordable in many major cities around the world, a significant portion of the population of the metropolises prefer to live in more affordable smaller satellite cities and to commute on daily basis to the central business district of the metropolis or to the industrial precinct. In the regional areas, the same process happens to the villages and the agro‐food business centers. The villages and the small regional cities are expanded and form the regiopolises. In both the metropolis and the regiopolis, there is a central business district which is the main commercial, cultural, political, and international center of the urban area.

Management of these large urban areas with populations in the scale of tens of millions is a major challenge. The energy consumption management, traffic control, planning various services, crime control, and many other essential services are impossible to provide efficiently in this scale. To manage these large urban environments, the sensor and camera networks as well as many other automatic data generation tools are used to provide insight about the megacities. The analysis of the data streams created automatically by this network is a challenging problem. There has been a significant body of research on the application of big data analytics for smart environment management. Yu et al. [1] proposed a decentralized data assessment framework using Blockchain for smart city management. Zhou et al. [2] investigated big data for health services of the smart city. Nazerdeylami et al. [3] proposed a smart coastline management system using deep neural networks. Abbasi et al. [4] proposed a deep privacy‐preserving framework for Internet of Robotic Things (IoRT) in the smart city. Zenkert et al. [5] proposed a new dataset for smart city mobility management. Yadav et al. [6] proposed a framework for using Internet of Things (IoT) and big data for smart city management. Xu et al. [7] proposed a framework for smart tourism using IoT and big data. Wang et al. [8] designed a framework for monitoring the conditions of the infrastructures of the smart city using big data. Shao et al. [9] proposed a framework for using the video surveillance cameras in the smart city for decision support. Muhammed et al. [10] proposed a new healthcare network for the smart city. Huang et al. [11] designed a sustainable smart city management system by adhering to the quality of service standards. Norouzi et al. [12] presented a new solution for green software‐defined networking (SDN) for the smart networks.

In the next decade, gradually the intelligent autonomous vehicles are replacing the traditional manually driven cars. Therefore, the concept of smart highway and road management is of significant importance in the next years. In the smart highway system, the intelligent vehicles can recognize each other and to communicate in the vehicular Adhoc networks. However, during the transition period from traditional to intelligent highways, the manually driven cars and the intelligent vehicles drive side by side on the roads. During this transition period as the current generation vehicle does not have electronic tagging and ID systems, it is necessary for the smart vehicles to recognize the old vehicles using the traditional method of Automatic License Plate Recognition (ALPR).

ALPR has been studied extensively in the past decades. Some of the proposed methods are relying on computer vision techniques. Most of the works in license plate recognition using semantic segmentation is based on binary classification which is easier compared with multi‐class classification. In the binary classification, the license plates are classified into the digits and the background. Abedin et al. [13] used contour properties and deep neural networks for license plate recognition. Abinaya et al. [14] proposed a morphological transform to rectify blurred license plate images in moving cars. Ahn et al. [15] proposed an algorithm for license plate area detection. Fomani et al. [16] presented a method for detecting license plates using adaptive morphological operations and adaptive thresholding. Ghoneim et al. [17] proposed using super‐resolution enhancement for improved license plate recognition.

Computer vision and machine learning methods are powerful tools for automating complex tasks and in recent years, the accuracy of these methods are increased significantly. Recently, deep learning resulted in very good accuracies in various complex pattern recognition problems [1821]. Image understanding in computer vision and machine learning passed an evolutionary route from finding just a single label for the whole scene and localizing objects to segmentation of every pixel in a dense image. One of the first deep learning‐based models proposed for segmenting an image was by Long et al. [22] which substitute the fully connected layer in the last layer of the deep neural network with a Softmax layer to assign labels to each pixel. However, these models were not perfect for understanding global information from the images. Therefore, the models have been improved by a series of post‐processing techniques like Conditional Random Field (CRF) proposed by Lafferty et al. [23], use of dilated convolutions to expand the size of filters [24], and recently introduced ReSeg network which leverages recurrent neural network (RNN) to help better fine‐tune the feature maps [25].

In this chapter, the parametric model is chosen for multi‐class ALPR from the low altitude angle suitable for the vision system of smart vehicles. In the proposed model, each digit is segmented using a deep neural network. The proposed algorithm is divided into two stages: (i) a model was trained for detection of the bounding box of the license plate, i.e., localization. In this stage, 500 images of cars in different environments are used for training the model. (ii) A model is trained to segment each digit in bounding boxes that computed during the first section. In this stage, a naive sample generator is used, which simultaneously generates two other outputs: ground truth for L2 loss and ground truth for cross‐entropy loss.

The rest of this chapter is organized as follows. The proposed deep ALPR framework is detailed in Section 10.2. The experimental design and the simulation scenarios are discussed in Section 10.3. Section 10.4 discusses the practical implementation of enterprise‐level big data analytics for smart internet of roads. Finally, Section 10.5 concludes the chapter.

10.2 Fully Convolutional Deep Neural Network for Autonomous Vehicle Identification

The proposed framework for license plate recognition consists of several steps which will be discussed in detail in this section.

10.2.1 Detection of the Bounding Box of the License Plate

A real‐life recognition application must pass two steps: detection and recognition. In this work, for detecting the license plates in the images, a convolutional neural network (CNN) is trained for semantic segmentation. The objective of this network is to find the bounding box of the license plates. The neural network is trained on roughly 500 images of cars in different environments. The output of the network is a vector containing the coordinates for the top‐left location of the bounding box of the license plate and the height and width of the plate. The CNN consists of three convolution layers and three pooling layers ending with a fully connected layer. The loss is an L2 loss that compares the output vector with the ground truth.

10.2.2 Segmentation Objective

When using a variational auto‐encoder (VAE), an objective function that will be used is an L2 loss function. Using an L2 loss objective generally seems to be a right choice in mapping or translation in which input and output have the same statistics, but different styles. However, in semantic segmentation, there is a better choice that is called Sparse Softmax Cross‐Entropy Loss which computes the cross‐entropy between Logits and labels. For using this objective function, we need another ground truth that has one band, and its values are constant and in the range of category labels.

10.2.3 Spatial Invariances

Pure CNN‐based models generally fail to capture all the invariances of the samples in semantic segmentation settings and because of this shortcoming, refinements are needed to solve this issue in the post‐processing or convergence stages [26]. Among these methods, CRF tries to keep the connection between low‐level image’s information and final segmented image. This technique helps to better understand the global information in samples and therefore segmenting objects with more details.

Furthermore, semantic segmentation models are highly sensitive to spatial invariances in the input image and a small change in input distribution could reduce segmentation accuracy. Therefore, in the proposed license plate segmentation method, all the spatial invariances are removed. In the pre‐processing stage, the license plates are centered in test samples. Removing spatial invariance means that the license plate centroid in images is at a constant location. The invariance strongly affects the generalization of model and forces it to perform worse on the unseen samples. Figure 10.1 shows the procedure of removing spatial invariances to improve the generalization and accuracy of semantic segmentation model.

10.2.4 Model Framework

10.2.4.1 Increasing the Layer of Transformation

Many segmentation models rely on Encoder–Decoder patterns without using any transformation stage. In VAE architectures, input samples must pass three stages. First, the given sample must be encoded through layers. Then, the result of encoding stage which has a lower dimensionality passes through transformation layers. This is the stage where all the transformation of input to output will happen. Finally, the decoding stage increases the dimension. In semantic segmentation, different numbers of layers in these stages affect the accuracy and results of the models. In our experiment, we saw that by increasing the number of layers in the transformation stage, considerable improvement in segmentation would occur. However, increasing the encoding and decoding layers would not necessarily lead to better‐segmented objects and details. Figure 10.2 shows the proposed architecture according to encoding‐transformation‐decoding network.

Photo depicts the procedure of removing spatial invariances to improve the generalization and the accuracy of semantic segmentation model.

Figure 10.1 The procedure of removing spatial invariances to improve the generalization and the accuracy of semantic segmentation model.

Schematic illustration of the architecture according to encoding-transformation-decoding network.

Figure 10.2 The architecture according to encoding‐transformation‐decoding network.

10.2.4.2 Data Format of Sample Images

In some applications of machine learning like object classification and object detection, feeding compressed format of input samples to the model helps to reduce the overfitting problem. However, in semantic segmentation, we avoided compression in ground truth for L2 loss to prevent losing the meaningful values.

10.2.4.3 Applying Batch Normalization

In this work, all the layers of the output are normalized using batch normalization with momentum of 0.5.

10.2.4.4 Network Architecture

A fully convolutional network [27] is used for semantic segmentation of license plates. The number of layers in encoding and decoding sections of the network is eight and three, respectively. In the transformation stage, four layers are used instead of one layer for better performance. Each layer uses leaky ReLU as the activation function and dropout with keep probability of 0.5. Sparse Softmax loss is chosen as an objective function. The optimization function is Adam optimizer which uses the learning rate of 0.0008 and beta of 0.5.

A deep neural network with convolution kernels is adopted as the core and primary architecture of this implementation. Figure 10.3 depicts the FCNN architecture. In FCNN, the convolution kernel produces either a downsized sample or performs up‐sampling using the transposed convolution (de‐convolution) kernels. Then, sequences of convolutions feed the learned data to a fully connected neural network for labeling.

10.2.5 Role of Data

The main purpose of using convolution on the samples is to extract features from them and based on these features find a meaningful correlation in data distributions. This is a more complicated procedure in objects like license plates where the aim is to segment each digit. This difficulty is because digits share a common texture with the background and the only correlation is the spatial location of digits. A workaround for this shortcoming is to use more data to learn the features. In our experiments, we compared the amount of data used and show that by having more data, the model has better performance in segmenting digits.

10.2.6 Synthesizing Samples

An obstacle in end‐to‐end learning models like deep neural networks is the preparation and pre‐processing of training samples. This problem is more evident in applications like semantic segmentation. In semantic segmentation, depending on the objective function, we could have up to two or three sets of samples. For instance, by selecting sparse Softmax loss, besides the training sample which is our real RGB image input (Width, Height, 3), we need another input as ground truth with shape of (Width, Height, 1). In this input, every pixel is labeled according to the list of category labels. However, if we decide to use L2 loss function, we need another ground truth input with shape of (Width, Height, 3) as depicted in Figure 10.4.

Schematic illustration of the architecture of FCNN.

Figure 10.3 The architecture of FCNN.

To resolve the difficulty in preparing samples, we decided to generate manually generated samples. In the process of sample generation, we generate a random number and a code and paste it in the license plate image. Also, two ground truth inputs are generated. Although this generated samples miss the exact details of the real counterparts, this process decreases the pre‐processing time. These generated samples are presented in Figure 10.5. Although these generated samples help the model to learn more from the data, they cannot replace the real data.

10.2.7 Invariances

There are many invariances that can affect the decision of an ALPR system and make it to produce false answers when facing an outlier. By adding invariants like illumination and blurring, we can hope that our model makes a more sensible decision. However, adding many spatial invariances could make the model useless. Therefore, the only spatial invariance that added in this research is a slight rotation. Furthermore, because clarity of the background in every car's license plate is different, in data generation, we randomly changed the background opacity to include this kind of invariances.

10.2.8 Reducing Number of Features

In the training stage, our models are dealing with a broad range of features. Sometimes, many of these features not only are useless for training a model, but they make the model worse. Therefore, in license plate segmentation where our objective is to only segment digits, the surrounding area of the license plate could not be considered as a useful source of features. Thus, we replace the background with a solid color.

Schematic illustration of ground truths needed for both types of objectives. (a) Sparse Softmax loss and (b) L2 loss.

Figure 10.4 Ground truths needed for both types of objectives. (a) Sparse Softmax loss and (b) L2 loss.

Schematic illustration of generated samples that used to train the segmentation model.

Figure 10.5 Generated samples that used to train the segmentation model.

10.2.9 Choosing Number of Classes

In multi‐class ALPR methods, the license plate is segmented into N regions. Usually, N is considered as the number of digits in license. In this work, we consider N as 5. Figure 10.3b shows the results of this segmentation.

Schematic illustration of segmented samples produced by the trained model.

Figure 10.6 Segmented samples produced by the trained model.

10.3 Experimental Setup and Results

For a more detailed assessment of applicability of the proposed segmentation model, we first present the Softmax loss of the proposed model on unseen samples during training and then compute the mean intersection over union (M‐IOU). This measure is extensively been used for semantic image segmentation assessment. Figure 10.6 presents the segmented samples produced by the proposed model after 70 iterations.

10.3.1 Sparse Softmax Loss

Table 10.1 shows the proposed model Softmax loss at final iteration.

Table 10.1 The proposed model Softmax loss at final iteration.

Dataset Sparse Softmax
Real data sample 0.5689

Table 10.2 Representation of the proposed model M‐IoU at final iteration.

Dataset M‐IoU
Unseen generated sample 0.87194
Real data sample 0.51665

10.3.2 Mean Intersection Over Union

M‐IoU, which is also referred to as Jaccard index, is a better evaluation method for our application. The objective of M‐IoU is to first compute the intersection of ground truth with the segmented output divided by union of each class and then averaging it over all existing classes. Table 10.2 presents the M‐IoU of learned model when applied to 100 of unseen and real data samples.

(10.1)equation

10.4 Practical Implementation of Enterprise‐Level Big Data Analytics for Smart City

In the real situation, the proposed ALPR for smart road or highway generates a large data stream of license plate numbers. The machine‐generated data and log files generated by the various components of the smart city sensor network and servers play an essential role in management of the smart city. These data and log files allow the system administrators to gain knowledge of operations, executions, and security issues regarding various smart city services. Significant increase in the volume of data and log files makes manual interpretation of these files very hard and in some cases even impossible. There are various machine learning and analytics tools for autonomous interpretation of log files for optimal management of services. These tools analyze very large log files and extract the required information for service administrators from these files autonomously.

The solution provided for analyzing this vast amount of data is distributed and parallel processing. The MapReduce‐based processing allows the analytics framework to split the data into smaller fractions and after processing these smaller datasets, the results are combined for the final insight into the data. The Hadoop‐based data processing frameworks are one of the solutions proposed for distributed processing of the big data. A very important step in smart log analysis is searching massive machine‐generated datasets. Due to the fact that searching large datasets and logs is very time consuming and usually the results of the analytics are required in real time, breaking the search into smaller searches is a vital ability for analyzing machine‐generated data.

A solution for this problem is using a combination of Indexers and Forwarders. The Indexers parse the machine‐generated data and the Forwarders deliver the data to the Indexers. To process large log streams in real time, the Indexing layer is distributed over several physical servers. In this distributed search environment, a Search Head has the responsibility to forward the query required by the system to parallel Indexers and each Indexer indexes a small batch of data. The Search Head also has the responsibility of aggregation of the results from the Indexers and to produce the final result. This process is demonstrated in Figure 10.7.

Schematic illustration of the Forwarder and Indexer architecture for parallel processing of smart city data.

Figure 10.7 The Forwarder and Indexer architecture for parallel processing of smart city data.

Schematic illustration of the Splunk front end and back end architecture for smart city data analytics.

Figure 10.8 The Splunk front end and back end architecture for smart city data analytics.

A necessary requirement for processing the smart city log files is consideration of timestamps of the data. The stream of sensor logs should be considered as a time series and only in this context, many events in the city has meaning. Therefore, to extract meaningful patterns from the smart city data streams, besides the physical parallelism of the indexing process, the data can be also parallel‐processed in time. The parallelism of indexing on data size refers to as Spatial MapReduce and the parallelism in time is called Temporal MapReduce. This solution is proposed by the Splunk for enterprise‐level data analytics. The back end of the Splunk uses massively parallel searching and indexing for finding patterns in the big data. The front end can be linked to various machine learning, visualization, and data analytics platforms for decision support for the smart city. The architecture of the Splunk front and back ends is demonstrated in Figure 10.8.

Schematic illustration of the clustering results for the traffic in Dublin.

Figure 10.9 The clustering results for the traffic in Dublin.

Schematic illustration of the classification results for the air pollution hotspots in Dublin.

Figure 10.10 The classification results for the air pollution hotspots in Dublin.

As a case study, a dataset from the city of Dublin, Ireland, is used. Two main scenarios are considered for evaluation of the big data analytics for smart city management. These scenarios are:

  • Service planning based on the commuter population.
  • Air pollution management.

The scenarios are investigated using the traffic in the city of Dublin. Clustering the traffic shows the sections which have highest traffic and therefore require optimized management. The output of this scenario is presented in Figure 10.9.

The next scenario is the classification results for the air pollution hotspots in Dublin. The output of this scenario is presented in Figure 10.10.

One of the most important issues in providing the results of the data analytics to the smart city management is visualization and dashboards. The dashboards should simulate the city map and offer insights to the management using the sensor network data. The large volume of data gathered by the sensor network of the smart metropolis should be presented in a comprehensible format for the human managers. A map‐based dashboard is necessary to address this issue.

10.5 Conclusion

In this chapter, an enterprise‐level big data analytics framework for autonomous management of road traffic in the smart cities is presented. The proposed framework is capable of providing decision support using large data streams. A deep neural network is used for segmentation of vehicle license plates in a complex and multi‐language environment. The performance of the proposed algorithm is evaluated using a dataset of real and manually generated data. The experimental results show that the proposed framework can detect and segment the license plates in complex scenarios and the results can be used in smart road applications.

References

  1. 1 Yu, H., Yang, Z., and Sinnott, R.O. (2018). Decentralized big data auditing for smart city environments leveraging blockchain technology. IEEE Access 7: 6288–6296.
  2. 2 Zhou, M. and Chen, X. (2018). Application of dendrobium compound preparation in the health environment of big data smart city. Paper presented at the 2018 International Conference on Intelligent Transportation, Big Data & Smart City (ICITBS), Xiamen, China.
  3. 3 Nazerdeylami, A., Majidi, B., and Movaghar, A. (2019). Smart coastline environment management using deep detection of manmade pollution and hazards. Paper presented at the 2019 5th Conference on Knowledge Based Engineering and Innovation (KBEI), Tehran, Iran.
  4. 4 Abbasi, M.H., Majidi, B., Eshghi, M., and Abbasi, E.H. (2019). Deep visual privacy preserving for internet of robotic things. Paper presented at the 2019 5th Conference on Knowledge Based Engineering and Innovation (KBEI), Tehran, Iran.
  5. 5 Zenkert, J., Dornhofer, M., Weber, C., Ngoukam, C., and Fathi, M. (2018). Big data analytics in smart mobility: modeling and analysis of the Aarhus smart city dataset. Paper presented at the 2018 IEEE Industrial Cyber‐Physical Systems (ICPS), St. Petersburg, Russia.
  6. 6 Yadav, P., and Vishwakarma, S. (2018). Application of Internet of Things and big data towards a smart city. Paper presented at the 2018 3rd International Conference On Internet of Things: Smart Innovation and Usages (IoT‐SIU), Bhimtal, India.
  7. 7 Xu, C., Huang, X., Zhu, J., and Zhang, K. (2018). Research on the construction of sanya smart tourism city based on internet and big data. Paper presented at the 2018 International Conference on Intelligent Transportation, Big Data & Smart City (ICITBS), Xiamen, China.
  8. 8 Wang, T., Bhuiyan, M.Z.A., Wang, G. et al. (2018). Big data reduction for a smart city's critical infrastructural health monitoring. IEEE Communications Magazine 56 (3): 128–133.
  9. 9 Shao, Z., Cai, J., and Wang, Z. (2017). Smart monitoring cameras driven intelligent processing to big surveillance video data. IEEE Transactions on Big Data 4 (1): 105–116.
  10. 10 Muhammed, T., Mehmood, R., Albeshri, A., and Katib, I. (2018). UbeHealth: a personalized ubiquitous cloud and edge‐enabled networked healthcare system for smart cities. IEEE Access 6: 32258–32285.
  11. 11 Huang, J., Xing, C.‐C., Shin, S.Y. et al. (2017). Optimizing M2M communications and quality of services in the IoT for sustainable smart cities. IEEE Transactions on Sustainable Computing 3 (1): 4–15.
  12. 12 Norouzi, A., Majidi, B., and Movaghar, A. (2018). Reliable and energy‐efficient routing for green software defined networking. Paper presented at the 2018 9th International Symposium on Telecommunications (IST), Tehran, Iran.
  13. 13 Abedin, M.Z., Nath, A.C., Dhar, P., Deb, K., and Hossain, M.S. (2017). License plate recognition system based on contour properties and deep learning model. Paper presented at the 2017 IEEE Region 10 Humanitarian Technology Conference (R10‐HTC), Dhaka, Bangladesh.
  14. 14 Abinaya, G., Banumathi, R., Seshasri, V., and Kumar, A.N. (2017). Rectify the blurred license plate image from fast moving vehicles using morphological process. Paper presented at the 2017 IEEE International Conference on Electrical, Instrumentation and Communication Engineering (ICEICE), Karur, India.
  15. 15 Ahn, C.S., Lee, B.‐G., Yang, S.‐S., and Park, S.‐C. (2017). Design of car license plate area detection algorithm for enhanced recognition plate. Paper presented at the 2017 4th International Conference on Computer Applications and Information Processing Technology (CAIPT), Kuta Bali, Indonesia.
  16. 16 Fomani, B.A. and Shahbahrami, A. (2017). License plate detection using adaptive morphological closing and local adaptive thresholding. Paper presented at the 2017 3rd International Conference on Pattern Recognition and Image Analysis (IPRIA), Shahrekord, Iran.
  17. 17 Ghoneim, M., Rehan, M., and Othman, H. (2017). Using super resolution to enhance license plates recognition accuracy. Paper presented at the 2017 12th International Conference on Computer Engineering and Systems (ICCES), Cairo, Egypt.
  18. 18 Fadaeddini, A., Majidi, B., and Eshghi, M. (2019). Privacy preserved decentralized deep learning: a blockchain based solution for secure ai‐driven enterprise. Paper presented at the International Congress on High‐Performance Computing and Big Data Analysis, Tehran, Iran.
  19. 19 Anbari, S., Majidi, B., and Movaghar, A. (2019). 3D modeling of urban environment for efficient renewable energy production in the smart city. Paper presented at the 2019 7th Iranian Joint Congress on Fuzzy and Intelligent Systems (CFIS), Bojnord, Iran.
  20. 20 Majidi, B., Patra, J.C., and Zheng, J. (2014). Modular interpretation of low altitude aerial images of non‐urban environment. Digital Signal Processing 26: 127–141.
  21. 21 Sanaei, S., Majidi, B., and Akhtarkavan, E. (2018). Deep multisensor dashboard for composition layer of web of things in the smart city. Paper presented at the 2018 9th International Symposium on Telecommunications (IST), Tehran, Iran.
  22. 22 Long, J., Shelhamer, E., and Darrell, T. (2015). Fully convolutional networks for semantic segmentation. Paper presented at the Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, USA.
  23. 23 Lafferty, J., McCallum, A., and Pereira, F.C. (2001). Conditional random fields: probabilistic models for segmenting and labeling sequence data. Paper presented at the Proceedings of the Eighteenth International Conference on Machine Learning, Williamstown, USA.
  24. 24 Chen, L.‐C., Papandreou, G., Kokkinos, I. et al. (2017). Deeplab: semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis and Machine Intelligence 40 (4): 834–848.
  25. 25 Visin, F., Ciccone, M., Romero, A., et al. (2016). Reseg: a recurrent neural network‐based model for semantic segmentation. Paper presented at the Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, Las Vegas, USA.
  26. 26 Garcia‐Garcia, A., Orts‐Escolano, S., Oprea, S. et al. (2018). A survey on deep learning techniques for image and video semantic segmentation. Applied Soft Computing 70: 41–65.
  27. 27 Shelhamer, E., Long, J., and Darrell, T. (2017). Fully convolutional networks for semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 39 (4): 640–651.
..................Content has been hidden....................

You can't read the all page of ebook, please click here login for view all page.
Reset
3.238.226.167