Automated Steel Bar Counting and Center Localization with Convolutional Neural Networks

1 Automated Steel Bar Counting and Center Localization with Con v olutional Neural Networks Zhun Fan, Senior Member , IEEE, Jie wei Lu, Benzhang Qiu, T ao Jiang, Kang An, Ale x Noel Josephraj, and Chuliang W ei Abstract —A utomated steel bar counting and center localization plays an important role in the factory automation of steel bars. T raditional methods only focus on steel bar counting and their performances ar e often limited by complex industrial en vironments. Convolutional neural network (CNN), which has great capability to deal with complex tasks in challenging en vi- ronments, is applied in this work. A framework called CNN-DC is proposed to achieve automated steel bar counting and center localization simultaneously . The proposed framework CNN-DC first detects the candidate center points with a deep CNN. Then an effective clustering algorithm named as Distance Clustering (DC) is proposed to cluster the candidate center points and locate the true centers of steel bars. The proposed CNN-DC can achieve 99.26% accuracy for steel bar counting and 4.1% center offset for center localization on the established steel bar dataset, which demonstrates that the proposed CNN-DC can perf orm well on au- tomated steel bar counting and center localization. Code is made publicly av ailable at: https://github.com/BenzhangQiu/Steel-bar - Detection. Index T erms —Steel bar counting, center localization, conv olu- tional neural network, distance clustering. I . I N T RO D U C T I O N S TEEL industry is one of the most important basic indus- tries for many countries. Steel bar is one of the most commonly used steel products, which is mainly used for building construction. Factory automation [1], [2], [3], [4], [5], [6] plays an essential role in impro ving the producti vity of steel bars, in which automated steel bar counting and center localization are among the most crucial steps. A. Steel Bar Counting Steel bar counting plays an important role in the manage- ment of steel production. T raditional steel bar counting is based on human calculation. Skilled workers count the number of steel bars in the factory , which is time-consuming and prone to errors. In recent years, some image processing techniques are employed to achiev e automated steel bar counting. In [7], Zhang et.al. use a template matching algorithm and a mutativ e threshold segmentation method to achie ve steel bar counting. In [8], Y ing et.al. combine Sobel operator and Hough transformation for automatic steel bar counting. In [9], Corresponding author: Chuliang W ei (email: clwei@stu.edu.cn). Jiewei Lu and Benzhang Qiu contributed equally . Authors are with the Guangdong Provincial Key Laboratory of Digital Signal and Image Processing, College of Engineering, Shantou University , Shantou 515063, China. This research work was supported by Guangdong K ey Laboratory of Digital Signal and Image Processing, the National Natural Science Foundation of China under Grant (61175073, 61300159, 61332002, 51375287). W u et.al . propose a steel bar counting method which utilizes concav e dots matching, K-level fault tolerance and visual feedback. In [10], Ghazali et.al. employ Hough transformation and the Laplacian of Gaussian (LoG) technique to perform automated steel bar counting. In [11], Liu et.al . propose a contours-based steel bar identification algorithm to count the number of steel bars. Ho wever , these methods are easily affected by the noisy and complex industrial en vironment, and need to adjust many parameters during operations. B. Center Localization Center localization means locating the center of each steel bar , as shown in Fig. 2. The reason of performing center localization is that it is an important step to achieve automated nameplate welding. Before leaving steel factories, steel bars need to be welded nameplates which contain the information of steel bars, such as production time and type specification. T raditional nameplate welding is performed by skilled work- ers. In the process of automated production, manipulators are used to weld nameplates on the steel bars, as shown in Fig. 3. Before a nameplate on a steel bar is welded, an appropriate welding point which is close to the center of a steel bar needs to be chosen in order to a void the broken and sliding problems of nameplates in transit. Therefore, locating the centers of steel bars is the first and one of the most crucial steps to achieve automated nameplate welding. In this paper , a framework called CNN-DC is proposed to achiev e automated steel bar counting and center localization simultaneously . The proposed CNN-DC frame work first de- tects the candidate center points with a deep conv olutional neural network (CNN). Then an effecti ve clustering algorithm called Distance Clustering (DC) is proposed to cluster the candidate center points and obtain the centers of steel bars. The experimental results demonstrate the ef fectiv eness of CNN-DC on steel bar counting and center localization. The rest of this paper is structured as follows: Section II describes the proposed CNN-DC framework. Section III introduces the steel bar dataset and e valuation metrics. Section IV provides the experimental results. Section V presents the conclusions of this paper . I I . M E T H O D O L O G Y The proposed CNN-DC framework can be regarded as a two-stage algorithm for automated steel bar counting and center localization. CNN-DC first obtains the candidate center points of steel bars with a deep CNN, and then an effecti ve 2 Fig. 1. The proposed CNN-DC framework. CNN-DC first detects the center points with a deep CNN, and then applies Distance Clustering algorithm to obtain the information of the number of steel bars and the center locations of the steel bars. Fig. 2. An example of center localization. Red points are the centers of steel bars. Fig. 3. Manipulators are used to weld nameplates on the steel bars. clustering algorithm named as Distance Clustering is proposed to cluster the candidate center points. Fig. 1 shows the frame- work of CNN-DC. A. Convolutional Neural Networks Con volutional neural networks are applied in our work to detect the candidate center points. Neural networks are inspired by biological processes [12], [13], and can be used to process a variety of high dimensional data [14], [15], [16], [17], [18], [19], [20], [21], [22], such as images, videos, voice signals and text characters. When dealing with the abov e- mentioned data, the application of fully connected networks is sometimes cumbersome due to its large feature space. Therefore, CNNs as special types of neural networks are preferred due to some important characteristics, such as spatial arrangement, sparse interactions, parameter sharing [23]. The input of CNNs is called tensor , which comprises of a multi-dimensional array . The core components of CNNs are con volutional and pooling layers. A con volutional layer con volv es the input tensor with a set of kernels to generate the output tensor . For each kernel, a feature map is generated by performing con volution, which slides the kernel on the whole spatial positions of the input tensor . Each con volutional layer consists of a set of kernels and thus produces a collection of feature maps, which are stacked together to generate the output tensor . When the input tensor is shaped as gre yscale image, the following steps are performed in order to specify the con volutional layer: (1) Accepting the input tensor with size H i × W i × D i , and obtaining the four parameters: K (the num- ber of kernels), F (the spatial dimensions of kernels), S (stride) and P (the zero padding size). (2) Producing the output tensor with size H o × W o × D o , where H o = ( H i − F + 2 P ) /S + 1 , W o = ( W i − F + 2 P ) /S +1 and D o = K . (3) The total number of parameters for the kernels is ( F × F × D i ) × K , where each kernel has ( F × F × D i ) parameters. Pooling layers are used to reduce the number of training parameters and control ov erfitting. They hav e two downsampling strategies: max- pooling and mean-pooling. Normally max-pooling is adopted with kernels of size 2 × 2 and stride 2 . Con volutional neural networks generally consist of sev eral con volutional layers and pooling layers, finalized with one or more fully connected layers. The details of CNNs used in this paper are provided in Fig. 5. B. Distance Clustering Clustering is a statistical analysis method applied to classification problems. The methods of clustering include: connectivity-based clustering [24], [25], centroid-based clus- tering [26], [27], density-based clustering [28] [29], grid- based clustering [30]. An effecti ve clustering algorithm called Distance Clustering (DC) is proposed to cluster the candidate center points obtained from the CNN. The pseudocode of DC is shown in Algorithm 1 and illustrated in Fig. 4: 3 Algorithm 1 : Distance Clustering (DC) Input: The locations of candidate center points and a distance threshold th d Output: A set of clusters, the center of each cluster and the number of clusters Step 1:Initialization 1) For i = 1 , . . . , n , set D ( i ) = d ij , where n is the number of candidate center points, d i is the Euclidean distance between the i th candidate center point and its closest candidate center point j , D is the set of d ij . 2) Create an empty structure S to sa ve the initial clusters. 3) For i = 1 , . . . , n , do if D(i) < th d , S { i } = { i;j } ; else S { i } = { i } end for Step 2:Clustering 1) Create a structure S c { 1 } = S { 1 } to save the final clusters and set the number of clusters as n S c = 1. 2) For i = 2 , . . . , n ,do n S c = length( S c ) condition = 0 For k = 1 , . . . , n S c ,do if ( S { i } ∩ S c { k } )! = ∅ , do S c { k } = { S c { k } ; S { i }} condition = 1 break end if end for if condition = 0, do n S c = n S c + 1 S c { n S c + 1 } = S { i }} end if end for Step 3:Obtaining Centers 1) Create an empty center set C to save the centers 2) For i = 1 , . . . , n S c ,do x c i = (max( x S c { i } ) + min( x S c { i } )) / 2 y c i = (max( y S c { i } ) + min( y S c { i } )) / 2 C ( i ) = ( x c i , y c i ) end for Initialization: In this step, for the i th candidate center point, its Euclidean distances with other candidate center points are calculated first. Then the closest distance d ij with the j th candidate center point is chosen and assigned to the i th candidate center point. Moreo ver , a structure S is created to sav e neighbors for the i th candidate center point. A distance threshold th d is used to decide whether the point j is a neighbor of the i th candidate center point and belongs to the set of S { i } . In the experiment, th d is set as 20. Clustering: A new structure S c is created to sav e the final clusters. For each clusters S { i } in S , if it has common elements with one of the clusters S c { k } in S c , the cluster S { i } will be merged into S c { k } . Otherwise, S c will create a new cluster to save S { i } . Obtaining Centers: For each cluster in S c , the localization of each cluster center is calculated by averaging the maximum and minimum coordinate values for the x-coordinate and y- coordinate. Fig. 4. Illustrating the process of distance clustering: (1) Calculating distances of each two candidate center points. (2) Creating a set for each candidate center point by distance threshold. (3) Merging sets if they have common elements. (4) Calculating the center point of each set. I I I . D A T A S E T A N D E V A L U A T I O N M E T R I C S A. Dataset The steel bar dataset consists of 10 images. The images are obtained by an industrial camera in a steel bar factory . Each image has a high resolution and is of size 1440 × 1080 . Moreov er, each image has RGB channels and each channel is with 8 bits. The 10 images were divided into a training set and test set. The training set contains 4 images while the test set consists of 6 images. 99195 patches extracted from the training set are used to train the network. B. Evaluation Metrics Four commonly used e valuation metrics are applied in our work to assess the performance of CNN-DC: Recal l = T P T P + F N P r ecision = T P T P + F P F 1 = 2 × P r ecision × Recal l P r ecision + Recal l where TP , FP and FN indicates true positi ve (the number of correctly detected center points), false positi ve (the number of incorrectly detected center points), and false negati ve (the number of undetected center points), respectiv ely . R ecall in- dicates the CNN-DC’ s ability of detecting center points while P r ecision is used to measure the CNN-DC’ s capability of correctly detecting center points. F 1 is a comprehensive index of Recal l and P r ecision . The calculation time of applying CNN-DC on each test image is also stored. In order to further ev aluate the performance of CNN-DC for automated steel bar counting and center localization, two other useful metrics are employed in this work. The first one is relativ e accuracy [31] : Acc r = (1 − | N d − N | N ) × 100% (1) 4 Fig. 5. The CNN architecture composed of four con volutional layers, four pooling layers and three fully connected layers. Layer names are followed by the number of feature maps. Square brackets specify the kernel size and stride. It is noted that ’con v’, ’maxpool’ and ’fc’ are short terms for conv olutional layer, max pooling layer and fully connected layer, respectively . Zero-padding is not used in this paper . Fig. 6. The patches with label 1. where N d = T P + F P is the number of detected center points. N is the actual number of steel bars. Acc r ev aluates the performance of CNN-DC on steel bar counting. The second is the offset degree of center points: of f set = Σ N i =1 X i m N × 100% (2) where m = 71 is the average diameter of a steel bar and used to perform data normalization, X i is the Euclidean distance between the i th manually marked center point and the closest detected center point. of f set is used to e valuate the performance of CNN-DC on center localization. I V . E X P E R I M E N T S In this section, the training setup of CNN is first introduced, followed by the experimental results. A. T raining Setup 1) Network Ar chitectur e and T raining P arameters: The network architecture used in our experiment is composed Fig. 7. The patches with label 0. of four con volutional layers, four pooling layers and three fully connected layers, as shown in Fig. 5. The network was trained by the stochastic gradient descent algorithm [32]. L 2 regularization with a weight decay 0.0001 was adopted to prev ent overfitting. The learning rate was set as 0.001 and the training was stopped after 40 epochs. The implementation of CNN-DC was based on T ensorflo w [33]. The training was conducted on a Intel Xeon E5-2690 CPU with a TIT AN Xp GPU. 2) Data Pr eparation: In our experiment, 4 steel bar images are used to train the network. The input of the network are patches extracted from images based on each image pixel. The patch size is of 71 × 71 ( 71 is approximately the diameter of the steel bars). 1440 × 1080 × 4 patches are extracted first. Then the patches whose centers were within the 7 × 7 rectangles centered in the manual center points are labeled as 1, and other patches are labeled as 0. Finally , 26468 patches are labeled as 1, while 6194332 patches are labeled as 0, which leads to an imbalanced data problem. In training, all patches with label 1 are chosen first. Then the patches with label 0 are selected randomly according to the ratio of positi ve (1) to negati ve (0) 5 Fig. 8. The process illustrating the application of CNN-DC: (1) A 71 × 71 sliding window is adopted to extracted patches from images based on each image pixel. (2) These patches are detected by CNN with architecture shown in Figure 5. (3) The result of CNN classification. (4) The center coordinates of the patches considered as label 1 are shown as red points, and the green boxes show the group of red points by distance clustering with process shown in Algorithm 1. (5) The center of each group is reimposed into the center of each steel bar of the original image accordingly . patches. In our experiment, the ratio of positiv e to negati ve patches is set as 1 : 3 . Thus 26468 patches with label 1 and 72727 patches with label 0 are used to train the network. The examples of positi ve and negati ve training examples are sho wn in Fig. 6 and 7. During testing, a sliding window with stride 6 was employed to improve the ef ficiency of CNN-DC. The process of selecting the optimal test stride with regard to the efficienc y of CNN-DC is provided in the third experiment of the next section. T ABLE I T H E P E RF O R MA N C E O F C N N - DC O N T H E S T EE L BA R D A T A S E T A verage Indexs of CNN-DC Recall Precision F1 0.9951 0.9976 0.9963 Acc r of f set times(s) 99.26% 4.11% 3.5862 B. Results Three e xperiments are conducted in order to demonstrate the effecti veness of the proposed CNN-DC on automated steel bar counting and center localization. In the first e xperiment, the performance of CNN-DC on steel bar counting and center localization was analyzed. In the second experiment, the proposed CNN-DC was compared with other methods. In the third experiment, the analysis of two parameters (the test stride and the distance threshold th d ) in DC algorithm was conducted. 1) The F irst Experiment: The performance of CNN-DC on the steel bar dataset is shown in T ABLE I. From T ABLE I, it can be observed that CNN-DC can obtain high scores on Recall, Precision and F1, with a verage v alues of 0 . 9951 , 0 . 9976 and 0 . 9963 , respecti vely , which indicates that CNN-DC can ef fectively identify center points on the steel bar images. Moreov er, CNN-DC can achiev e a high score on Acc r with an av erage value of 99 . 26% , which means that CNN-DC can ha ve a good performance in steel bar counting. Moreover , the low of f set score ( 4 . 1% ) indicates that CNN-DC performs well on center localization. In addition, the calculation time of CNN- DC ( time = 3 . 5862 s ) indicates that the CNN-DC can meet the requirement of real-time processing for f actory automation. The processes illustrating the application of CNN-DC on the steel bar dataset are shown in Fig. 8. 2) The Second Experiment: In the first experiment, it has been demonstrated that CNN-DC performs well on steel bar counting and center localization. In order to further demon- strate the effecti veness of CNN-DC, we compare CNN-DC with other existing methods on the steel bar dataset. From T ABLE II, it can be observed that CNN-DC outperforms other methods in terms of Recall, Precision, F1, Acc r and of f set . 6 Zhang et.al . [7] Y ing et.al . [8] Ghazali et.al . [10] Liu et.al . [11] Proposed The intermediate result The final result Fig. 9. Comparison of results obtained by different methods on an exemplar image. In Zhang et.al . [7], the template matching algorithm can not match the image which is not similar to the template, and fails to identify some steel bars. In Y ing et.al. [8] and Ghazali et.al. [10], the Hough transformation algorithm is sensitive to edge information. Some circular background areas and steel bars with blur edges are misidentified. In Liu et.al. [11], the contour-based algorithm relies on good contour extraction. It is sensitive to luminance v ariation of steel surface and edge blurring of steel bars. Some single steel bars are misidentified multiple steel bars, and Some multiple steel bars are misidentified single steel bars. The proposed method has better robustness for above disturbance, which performs well on en vironmental disturbance, luminance variation of steel surface and edge blurring of steel bars. 7 1 5 9 13 17 Stride 0.4 0.6 0.8 1 Value Recall Precision F1 1 5 9 13 17 Stride 0.4 0.6 0.8 1 Acc r (a) (b) 1 5 9 13 17 Stride 0 0.2 0.4 0.6 offset 1 5 9 13 17 Stride 0 100 200 300 time(s) (c) (d) Fig. 10. V ariations of dif ferent e valuation metrics on the training data of the steel bar dataset with the increase of stride. (a) The variations of Recall, Precision and F1 with the increase of stride. (b) The variation of Acc r with the increase of stride. (c) The variation of offset with the increase of stride. (d) The v ariation of calculation time with the increase of stride. 1 11 21 31 41 51 61 71 th d 0 0.5 1 Value Recall Precision F1 1 11 21 31 41 51 61 71 th d -3 -2 -1 0 1 Acc r (a) (b) 1 11 21 31 41 51 61 71 th d 0 1 2 3 offset 1 11 21 31 41 51 61 71 th d 2 4 6 time(s) (c) (d) Fig. 11. V ariations of different ev aluation metrics on the training data of the steel bar dataset with the increase of th d . (a) The variations of Recall, Precision and F1 with the increase of th d . (b) The variation of Acc r with the increase of th d . (c) The variation of offset with the increase of th d . (d) The variation of calculation time with the increase of th d . 8 T ABLE II T H E C O MPA R IS O N B E T WE E N C N N - DC A N D O T HE R M E T H OD S A verage Recall Precision F1 Acc r of f set times(s) Zhang et.al . [7] 0.8864 0.9360 0.9103 94.69% 15.83% 0.3023 Y ing et.al. [8] 0.9617 0.8417 0.8975 85.68% 12.58% 0.2404 Ghazali et.al . [10] 0.9778 0.9366 0.9566 95.56% 10.30% 0.1346 Liu et.al . [11] 0.8123 0.6833 0.7420 80.99% 25.58% 0.0313 Proposed 0.9951 0.9976 0.9963 99.26% 4.11 % 3.5862 As shown in Fig. 9, the result obtained by other methods is sensitiv e to en vironmental disturbance, luminance variation of steel surface and edge blurring of steel bars. The proposed method has better robustness for challenging en vironments. Although the calculation time of CNN-DC is higher than other methods, it can be reduced by model acceleration methods. With the help of network binarization[34], structured pruning[35] and matrix decomposition[36], the calculation time can be further reduced and will be the next step in our future work. 3) The Thir d Experiment: In our work, str ide = 6 and th d = 20 are used, which were selected based on the experiments on the training data of the steel bar dataset. In order to illustrate the reasons of choosing these values, the variations of the ev aluation metrics: Recall, Precision, F1, of f set and the calculation time on the training data of the steel bar dataset by varying str ide and th d are giv en in Fig. 10 and 11. From Fig. 10, it can be observed that CNN-DC can hav e a good performance on the training data of the steel bar dataset as str ide v aries in [1,9]. In the meantime, CNN- DC can hav e a low calculation time as str ide v aries in [5,18]. From the above observation, CNN-DC performs overall better as str ide varies in [5,9], and str ide = 6 is chosen randomly from [5,9]. From Fig. 11, it can be observed that CNN-DC can maintain high scores of Recall, Precision, F1 and Acc r as th d varies in [11,51]. CNN-DC can also have a low of f set v alue as th d varies in [1,51]. Moreov er, the calculation time varies very slightly with change of th d . From the above observ ation, CNN-DC can obtain a good performance as th d varies [11,51], and th d = 20 is selected randomly from [11,51]. V . C O N C L U S I O N S Automated steel bar counting and center localization are of great significance in factory automation of steel bars. Steel bar counting and center localization are traditionally performed by skilled workers, which are tedious and time-consuming. In order to alleviate the burdens of workers on steel bar counting and center localization, an effecti ve framew ork called CNN- DC is proposed to achiev e steel bar counting and center lo- calization simultaneously . The proposed CNN-DC framework first performs candidate center point detection with a deep con volutional neural network, which is followed by a Distance Clustering algorithm to cluster the candidate center points and obtain the center locations of steel bars. The experimental results show that the proposed CNN-DC framework performs well on steel bar counting and center localization, achieving Recal l = 0 . 9951 , P r ecision = 0 . 9976 , F 1 = 0 . 9963 , Acc r = 99 . 26% , of f set = 4 . 11% and time = 3 . 5862 s . In the future, more steel bar images will be collected by negotiating with the managers of steel bar factories in order to further validate the effecti veness of CNN-DC. Applying CNN-DC real time in steel bar factories is also planned in order to validate the practicability of the proposed method. R E F E R E N C E S [1] V . E. W agner , A. A. Andreshak, and J. P . Staniak, “Power quality and factory automation, ” IEEE Tr ansactions on Industry Applications , vol. 26, no. 4, pp. 620–626, 1990. [2] Y . Miyatake and R. Kangari, “Experiencing computer integrated con- struction, ” Journal of Construction Engineering and Management , vol. 119, no. 2, pp. 307–322, 1993. [3] H.-J. K orber, H. W attar, and G. Scholl, “Modular wireless real-time sensor/actuator network for factory automation applications, ” IEEE T ransactions on Industrial Informatics , vol. 3, no. 2, pp. 111–119, 2007. [4] A. Grau, M. Indri, L. L. Bello, and T . Sauter, “Industrial robotics in factory automation: From the early stage to the internet of things, ” in IECON 2017-43r d Annual Conference of the IEEE Industrial Electr onics Society . IEEE, 2017, pp. 6159–6164. [5] M. Dotoli, A. Fay , M. Mi ´ sko wicz, and C. Seatzu, “ Adv anced control in factory automation: a survey , ” International Journal of Production Resear ch , vol. 55, no. 5, pp. 1243–1259, 2017. [6] M. Dotoli, A. Fay , M. Miko wicz, and C. Seatzu, “ An overvie w of current technologies and emerging trends in factory automation, ” International Journal of Production Researc h , pp. 1–21, 2018. [7] D. Zhang, Z. Xie, and C. W ang, “Bar section image enhancement and positioning method in on-line steel bar counting and automatic separating system, ” in Image and Signal Pr ocessing, 2008. CISP’08. Congr ess on , vol. 2. IEEE, 2008, pp. 319–323. [8] X. Y ing, X. W ei, Y . Pei-xin, H. Qing-da, and C. Chang-hai, “Research on an automatic counting method for steel bars’ image, ” in Electrical and Contr ol Engineering (ICECE), 2010 International Confer ence on . IEEE, 2010, pp. 1644–1647. [9] Y . W u, X. Zhou, and Y . Zhang, “Steel bars counting and splitting method based on machine vision, ” in Cyber T echnology in Automation, Contr ol, and Intelligent Systems (CYBER), 2015 IEEE International Conference on . IEEE, 2015, pp. 420–425. [10] M. F . Ghazali, L.-K. W ong, and J. See, “ Automatic detection and counting of circular and rectangular steel bars, ” in 9th International Confer ence on Robotic, V ision, Signal Pr ocessing and P ower Applica- tions . Springer, 2017, pp. 199–207. [11] L. Xiaohu and O. Jineng, “Research on steel bar detection and count- ing method based on contours, ” in 2018 International Conference on Electr onics T echnology (ICET) . IEEE, 2018, pp. 294–297. [12] M. Matsugu, K. Mori, Y . Mitari, and Y . Kaneda, “Subject independent facial expression recognition with robust face detection using a conv olu- tional neural network, ” Neural Networks , vol. 16, no. 5-6, pp. 555–559, 2003. [13] Y . LeCun, L. Bottou, Y . Bengio, and P . Haffner , “Gradient-based learning applied to document recognition, ” Proceedings of the IEEE , vol. 86, no. 11, pp. 2278–2324, 1998. [14] O. Abdel-Hamid, A.-r. Mohamed, H. Jiang, L. Deng, G. Penn, and D. Y u, “Con volutional neural networks for speech recognition, ” IEEE/ACM T ransactions on Audio, Speech, and Language Pr ocessing , vol. 22, no. 10, pp. 1533–1545, 2014. [15] U. R. Acharya, S. L. Oh, Y . Hagiwara, J. H. T an, and H. Adeli, “Deep con volutional neural network for the automated detection and diagnosis of seizure using eeg signals, ” Computers in Biology and Medicine , vol. 100, pp. 270–278, 2018. 9 [16] B. Hu, Z. Lu, H. Li, and Q. Chen, “Con volutional neural network architectures for matching natural language sentences, ” in Advances in Neural Information Pr ocessing Systems , 2014, pp. 2042–2050. [17] S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: tow ards real- time object detection with re gion proposal networks, ” IEEE T ransactions on P attern Analysis & Machine Intelligence , no. 6, pp. 1137–1149, 2017. [18] H. Su, S. Maji, E. Kalogerakis, and E. Learned-Miller , “Multi-view con volutional neural networks for 3D shape recognition, ” in Proceedings of the IEEE International Conference on Computer V ision , 2015, pp. 945–953. [19] J. Sun, W . Cao, Z. Xu, and J. Ponce, “Learning a conv olutional neural network for non-uniform motion blur remov al, ” in Proceedings of the IEEE Confer ence on Computer V ision and P attern Recognition , 2015, pp. 769–777. [20] R. Sakurai, S. Y amane, and J.-H. Lee, “Restoring aspect ratio distortion of natural images with conv olutional neural network, ” IEEE Tr ansac- tions on Industrial Informatics , vol. 15, no. 1, pp. 563–571, 2019. [21] Z. Zheng, Y . Y ang, X. Niu, H.-N. Dai, and Y . Zhou, “W ide and deep con volutional neural networks for electricity-theft detection to secure smart grids, ” IEEE T ransactions on Industrial Informatics , vol. 14, no. 4, pp. 1606–1615, 2017. [22] H. Gao, B. Cheng, J. W ang, K. Li, J. Zhao, and D. Li, “Object classification using cnn-based fusion of vision and lidar in autonomous vehicle environment, ” IEEE Tr ansactions on Industrial Informatics , vol. 14, no. 9, pp. 4224–4231, 2018. [23] A. Krizhevsk y , I. Sutskever , and G. E. Hinton, “Imagenet classification with deep con volutional neural networks, ” in Advances in Neural Infor- mation Processing Systems , 2012, pp. 1097–1105. [24] T . Zhang, R. Ramakrishnan, and M. Livny , “Birch: an efficient data clustering method for very large databases, ” in ACM Sigmod Record , vol. 25, no. 2. ACM, 1996, pp. 103–114. [25] G. Karypis, E.-H. S. Han, and V . Kumar , “Chameleon: Hierarchical clustering using dynamic modeling, ” Computer , no. 8, pp. 68–75, 1999. [26] J. A. Hartigan and M. A. W ong, “ Algorithm as 136: A K-means clustering algorithm, ” Journal of the Royal Statistical Society . Series C (Applied Statistics) , vol. 28, no. 1, pp. 100–108, 1979. [27] R. T . Ng and J. Han, “Clarans: A method for clustering objects for spatial data mining, ” IEEE T ransactions on Knowledge & Data Engineering , no. 5, pp. 1003–1016, 2002. [28] M. Ester, H.-P . Kriegel, J. Sander, X. Xu et al. , “ A density-based algorithm for discovering clusters in large spatial databases with noise, ” in ACM SIGKDD Conference on Knowledge Discovery and Data Mining(KDD) , vol. 96, no. 34, 1996, pp. 226–231. [29] A. Rodriguez and A. Laio, “Clustering by fast search and find of density peaks, ” Science , vol. 344, no. 6191, pp. 1492–1496, 2014. [30] M. Ankerst, M. M. Breunig, H.-P . Kriegel, and J. Sander , “Optics: ordering points to identify the clustering structure, ” in A CM Sigmod Recor d , vol. 28, no. 2. A CM, 1999, pp. 49–60. [31] J. S. Armstrong and F . Collopy , “Error measures for generalizing about forecasting methods: Empirical comparisons, ” International journal of for ecasting , vol. 8, no. 1, pp. 69–80, 1992. [32] L. Bottou, “Stochastic gradient descent tricks, ” in Neural networks: T ricks of the trade . Springer, 2012, pp. 421–436. [33] M. Abadi, P . Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. De vin, S. Ghemaw at, G. Irving, M. Isard et al. , “T ensorflow: A system for lar ge- scale machine learning, ” in 12th { USENIX } Symposium on Operating Systems Design and Implementation ( { OSDI } 16) , 2016, pp. 265–283. [34] M. Rastegari, V . Ordonez, J. Redmon, and A. Farhadi, “Xnor-net: Imagenet classification using binary con volutional neural networks, ” in Eur opean Confer ence on Computer V ision . Springer , 2016, pp. 525– 542. [35] J.-H. Luo, J. W u, and W . Lin, “Thinet: A filter level pruning method for deep neural network compression, ” in Pr oceedings of the IEEE international conference on computer vision , 2017, pp. 5058–5066. [36] Y .-D. Kim, E. P ark, S. Y oo, T . Choi, L. Y ang, and D. Shin, “Compression of deep conv olutional neural networks for fast and low power mobile applications, ” arXiv pr eprint arXiv:1511.06530 , 2015.