https://orcid.org/0000-0002-7166-4440
Figures
Abstract
This paper proposes a method for classifying the river state (a flood risk exists or not) from river surveillance camera images by combining patch-based processing and a convolutional neural network (CNN). Although CNN needs much training data, the number of river surveillance camera images is limited because flood does not frequently occur. Also, river surveillance camera images include objects that are irrelevant to the flood risk. Therefore, the direct use of CNN may not work well for the river state classification. To overcome this limitation, this paper develops patch-based processing for adjusting CNN to the river state classification. By increasing training data via the patch segmentation of an image and selecting patches that are relevant to the river state, the adjustment of general CNNs to the river state classification becomes feasible. The proposed patch-based processing and CNN are developed independently. This yields the practical merits that any CNN can be used according to each user’s purposes, and the maintenance and improvement of each component of the whole system can be easily performed. In the experiment, river state classification is defined as the following problems using two datasets, to verify the effectiveness of the proposed method. First, river images from the public dataset called Places are classified to images with Muddy labels and images with Clear labels. Second, images from the river surveillance camera in Nagaoka City, Japan are classified to images captured when the government announced heavy rain or flood warning and the other images.
Citation: Oga T, Harakawa R, Minewaki S, Umeki Y, Matsuda Y, Iwahashi M (2020) River state classification combining patch-based processing and CNN. PLoS ONE 15(12): e0243073. https://doi.org/10.1371/journal.pone.0243073
Editor: Guy J-P. Schumann, Bristol University/Remote Sensing Solutions Inc., UNITED STATES
Received: April 17, 2020; Accepted: November 14, 2020; Published: December 3, 2020
Copyright: © 2020 Oga et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The data used in our study are available from http://places2.csail.mit.edu/index.html and https://www.kishou.city.nagaoka.niigata.jp/camera/guest/.
Funding: This research is partially supported by SECOM Science and Technology Foundation, and Adaptable and Seamless Technology transfer Program through Target-driven R&D (A-STEP) from Japan Science and Technology Agency (JST) Grant Number JPMJTM20DJ. There was no additional external funding received for this study.
Competing interests: The authors have declared that no competing interests exist.
Introduction
In Japan, there has been an increasing number of annual occurrences of heavy rainfall with hourly precipitation of more than 50 mm (https://www.data.jma.go.jp). Accordingly, flood damage has been increasing in many places. Therefore, we need to develop a method to prevent or reduce flood damage. To assess flood risk, it is necessary to monitor the appearance of a river.
To satisfy this requirement, methods using water level gauges have been proposed [1–3]. These methods detect floods by measuring the rise of the water level. Recently, water level gauges with ultrasonic technology [1] and those with a radar technique [2] have been used. These methods emit radio waves or ultrasonic waves and measure the time taken to return from the measured point on the water surface; from this information, the water level can be determined. With these techniques, we do not need to locate the water level gauges underwater and can avoid malfunctions caused by flowing water. However, these methods have the problem that they cannot accurately measure the water level if sediment exists on the measured points. There is also the problem that the cost of installing the equipment with large wireless communication sensors is high [4]. As another solution to monitoring river optical properties, a method using a turbidity meter has been proposed [5]. Because this method measures the muddiness underwater, the burden of installed the equipment is greater and it cannot be used when the water flow is rough. Furthermore, methods for mapping the flood inundation areas in satellite images have been proposed [6, 7]. However, these methods have high cost to capturing the satellite images with sensors.
To solve these problems, methods have been proposed for measuring the water level by capturing images of rivers [4, 8–11]. These methods have the merit that the cost of locating the equipment is low because only general surveillance cameras are needed. In fact, although water level gauges are not likely to be located in small rivers [4], surveillance cameras can be easily located in small rivers, as well as major rivers [12]. Therefore, by analyzing images obtained by a surveillance camera, it is possible to monitor the appearance of rivers without the need for special sensors such as water level gauges or turbidity meters. Specifically, a method using a surveillance camera has been proposed [13], which performs binary classification of whether a river appearance is normal or a potential flood risk, using a convolutional neural network (CNN) [14–20]. To improve the classification performance, this method performs image segmentation using a pyramid scene parsing network (PSPNet) [21], to remove irrelevant background areas in advance. To the best of our knowledge, this study [13] is the first work on classification of river appearance using a surveillance camera. However, this method has the following two problems. First, because there is a limited amount of training data when a flood risk exists, the performance may be poor. Second, because images include some areas that are irrelevant to classification, the irrelevant areas may decrease the performance.
In this paper, we propose a method for classifying the river state (a flood risk exists or not) from river surveillance camera images by combining patch-based processing and CNN. The patch-based processing is a technique for training CNN using multiple patches in an image [22]. When the number of training images for CNN is limited, the patch-based processing is beneficial for data augmentation. Specifically, we first produce more training data for flood risk by generating multiple patches from each image. In this processing, by generating multiple patches from each image, in a sliding window fashion [23], we can produce more training data for flood risk, to train the CNN. Thus, we can avoid the performance degradation that is caused by overfitting because of the small amount of training data. Second, we derive a patch selection scheme for obtaining only relevant patches. In this processing, we remove patches that are irrelevant to the assessment of flood risk; these are the patches (e.g., patches corresponding to the background area) that are not useful for the classification. Furthermore, we apply CNN to the selected patches and obtain the final classification results by majority voting [16, 24]. Specifically, the majority voting first produces classification results (of whether a flood risk exists), i.e., votes, for each patch in an image. Then, the majority voting classifies the image to the class receiving the largest number of votes. This enables the method to emphasize the effect of patches that are relevant to the flood risk, and successful classification (of whether a flood risk exists) becomes feasible.
Aim of this work
We aim at not development of a new kind of CNN but development of pre-processing and post-processing that can be introduced into any CNN. Recently, CNN software (e.g., Microsoft Azure (https://azure.microsoft.com/ja-jp/), IBM Watson (https://www.ibm.com/watson/jp-ja/), Amazon Web Services (https://aws.amazon.com/jp/) and Google Cloud (https://cloud.google.com/products/ai)) has become widespread. A user can introduce our patch-based processing into each user’s CNN. In the system development, it is often that multiple providers develop each component of the system independently and perform system integration finally. Because our patch-based processing and CNN are developed independently, the maintenance and improvement of each component of the whole system can be easily performed. For example, if a new CNN is presented, we can classify the river state by only replacing the current CNN with the new CNN without adjustment of the cooperation of CNN and patch-based processing. Experimental results with real river surveillance camera images show that the adjustment of general CNNs [14–20] to the river state classification becomes feasible.
Proposed pre-processing and post-processing for river state classification
Fig 1 shows an overview of the proposed method, which consists of the following five phases:
- image segmentation for obtaining water regions;
- patch-based data augmentation for CNN;
- selection of patches relevant to classification;
- training of CNN using the selected patches;
- testing with an ensemble of multiple patches for classification.
The details of each processing phase are shown below.
A. Image segmentation for obtaining water regions
In this process, we detect water regions in an image and perform zero padding for other regions, so that the subsequent processing can use only regions relevant to the classification. Specifically, we perform semantic segmentation based on PSPNet [21], which is currently one of the state-of-the-art methods. PSPNet achieved high performance for the ADE20K dataset [25], which consists of 150 classes (see Table 1), including classes related to water regions. In this study, we therefore trained PSPNet by using the ADE20K dataset. Because water regions may be classified as “water”, “sea”, “lake”, and “natatorium”, as well as “river”, we defined the regions classified as those classes to be water regions. In this way, we were able to successfully detect water regions from a river image.
Here, we explain the detailed structure of PSPNet. Many CNNs [14–20] consist of convolutional layers, pooling layers, and fully-connected layers. In contrast, some CNNs for semantic segmentation [21, 26, 27] replace fully-connected layers by convolutional layers and output a two-dimensional map instead of classes. Fig 2 illustrates the flow of PSPNet. First, we extract a feature map using a residual neural network (ResNet) [18], without fully-connected layers, as an encoder. We then obtain feature maps containing information with different scales by inputting a feature map to the pooling layers, which have various sizes. Through convolution of these feature maps, we can discover the global and local characteristics of an image. Furthermore, we perform upsampling to change each feature map to the size of an original feature map and concatenate the resulting maps; this produces a final feature map that possesses both global and local characteristics of an input image. Finally, we obtain the segmentation result by applying 1 × 1 convolution to the final feature map. As a result, PSPNet enables accurate segmentation of an image that includes water regions and various backgrounds.
B. Patch-based data augmentation for CNN
In this process, we perform data augmentation for CNN, i.e., patch generation from each river image. In general, features that are necessary for classification may be lost by dividing an image into patches. However, in our river state classification, we do not need to determine the detailed texture in local areas, because the global characteristics of the water regions are crucial for predicting flood risk. Therefore, we do not need to tackle this problem; it is sufficient to simply generate multiple patches from each river image. In the experiment shown below, we generated patches of 32 pixels × 32 pixels whose slide width is 16 pixels. The number of patches generated from a single image is defined as follows:
(1)
where N is the number of patches, I is the image width, P is the sliding window size, and C is the patch width, assuming that all images and patches are square. In this manner, we can increase the amount of data to compensate for insufficient of the training data for river state classification.
C. Selection of patches relevant to classification
In this process, we select patches that are relevant to flood risk from those generated for the training of the CNN. Specifically, by removing patches with zero-padded pixels, (i.e., non-water regions), we select only the relevant patches (i.e., water regions), which are then used to suitably train the CNN. For this purpose, we need to define the criterion for removing the patches. In the proposed method, we define Th as the maximum fraction of pixels whose values are zero; patches with a greater proportion of zero pixels are removed. We determine a suitable criterion by changing Th; we explain how Th was set in the experiment shown below.
D. Training of CNN using selected patches
In this process, the selected patches are used to train the CNN for the classification of the river state. Our proposal is not the design of a new CNN model but the development of a classification method that can incorporate any CNN. In the experiment, we demonstrate that the proposed method improves the performance of various types of CNNs, including AlexNet [14], NIN [15], GoogLeNet [16], VGGNet [17], ResNet [18], WideResNet [19], and ResNeXt [20].
E. Testing with an ensemble of multiple patches for classification
Here, we describe the test phase using the trained model for river state classification. When targeting river images, we can accurately classify patches showing water regions, but the performance of classification may be worse for non-water region patches. However, the proposed method can remove the non-water region patches. Therefore, a simple majority voting strategy, as used in [24], enables the successful classification of the river state.
Specifically, we generate multiple patches from each test image in the same manner as training. We then input the resulting patches to the trained CNN model and calculate the probability that a patch belongs to a water region. By using multiple classification results per image, we obtain the final classification result by majority voting. This results in accurate classification even though several patches may have been misclassified by the CNN.
Fig 3 shows an example of the testing phase of the proposed method. The middle of the figure illustrates image segmentation and patch-based data augmentation. The figure shows that these processes enable data augmentation when selecting relevant regions for classification. However, some irrelevant regions remain (such as the road regions in the bottom left of the image). Patch-based ensemble learning —that is, majority voting of the patch-wise test results— reduces the influence of the irrelevant regions, to realize accurate classification.
Experimental results
Experimental results with real river images are presented to verify the effectiveness of the proposed method. As described above, the purpose of this work is the development of patch-based processing to adjust general CNNs to the river state classification. We verify that our proposed processing satisfies this purpose below.
Dataset
The images were obtained from a dataset used for the previous studies [13, 28] and a dataset containing images from the river surveillance camera in Nagaoka City, Japan. The images in the former dataset are available from the database at http://places2.csail.mit.edu/index.html. We used images with a river label from this database. The images in the latter dataset are available from https://www.kishou.city.nagaoka.niigata.jp/camera/guest/. When flood risk increases, the color of the river appears to be brownish. Therefore, for the former dataset [13], we defined the river images with Muddy and Clear labels as positive and negative samples, respectively. The Muddy and Clear labels were defined in the previous study [13]. For the latter dataset, we defined the river images captured when the government announced heavy rain or flood warning as positive samples and other images as negative samples. The sizes of all images were 224 pixels × 224 pixels. We set the patch size and slide width as C = 32 and P = 16, respectively. Thus, by Eq (1), 169 patches were generated from each image. For training images, we performed contrast normalization [29] and mirroring [14]. Contrast normalization comprised both subtractive normalization and divisive normalization. Subtractive normalization is a process that subtracts the mean of all pixel values, of training data, from each pixel value [29]. This process normalizes the brightness of the input images, but may enhance noise. To tackle this problem, we also performed divisive normalization [29], which divides the results of subtractive normalization by the standard deviation of all pixel values of the training data.
Experimental conditions
In this experiment, we verify the effectiveness of the proposed method by applying our method to recently proposed CNNs [14–20] (see Table 2). We used ResNet models whis 20, 50, and 101 layers. The widening parameter for WideResNet, k, was set to 4. In the training of the CNN model, we trained a model by solving the optimization problem using binary cross entropy as the loss function. The loss function L is defined as follows:
(2)
where t is the ground truth class and y is the classification result from a model. For the solver of the optimization problem based on the loss function, we used stochastic gradient descent with momentum (momentum SGD) [30]. Momentum SGD [30] is an improved version of the SGD [31] algorithm, one of the most popular optimization algorithms used in deep learning. We set momentum and weight decay [32] to 0.9 and 0.0005, respectively. Here, weight decay is a scheme for avoiding overfitting caused by an excessive increase in weights. The number of epochs was set to 100. For AlexNet [14], NIN [15] VGGNet [17], and ResNet [18], we set the learning rate to 0.001, 0.001, 0.05, and 0.01, respectively; for other CNN models, we set the learning rate to 0.1. We multiplied the learning rate by 0.1 when the epoch number became 50. In addition, the mini-batch size was 32, and the initial values for the convolution filter were set as in [33]. Other parameters, such as the number of filters, were set in the same manner as in the original references.
Parameter setting
Here, we describe the setting of the parameter Th, which is used for selecting patches relevant to the classification. In this study, we validated the values Th = 20, 40, 60, 80, 100, which are expressed as percentages. We used 800 and 200 images for training and test, respectively. In this validation, we adopted AlexNet [14] as a CNN model. Fig 4 shows the result of validation; it shows that a more accurate result can be obtained with smaller value of Th. Therefore, we can conclude that Th should be small. If Th was set to less than 20%, all patches were removed, so that it was not possible to perform classification. Consequently, we defined Th as 20% for this experiment.
We describe a scheme for adjusting Th when we apply the proposed method to hundreds or thousands of river surveillance cameras. In such a situation, the suitable Th may be different depending on each photographing condition (e.g. places and the angle of view). Therefore, we should first collect images obtained in the similar conditions. Then, for each condition, we can adjust the suitable Th by dividing images into training and test images and validating the classification accuracy and Th.
Results
For the evaluation, we performed five-fold cross validation [34] and calculated the following F-measure:
(3)
where
Here, TP is the number of samples that are correctly classified as positive samples. FP is the number of samples that are misclassified as positive samples. FN is the number of samples that are misclassified as negative samples. TN is the number of samples that are correctly classified as negative samples.
Table 3 shows the classification results for river surveillance camera images. Here, the columns show the results of our method (ours), the previous method [13] (Baseline 1), and the original CNN (Baseline 2). By comparing ours with Baseline 2, we confirmed that the adjustment of general CNNs to the river state classification became feasible by our pre-processing and post-processing. Note that Baseline 1 is a method that tries to adjust the general CNNs to the river state classification, but this method only incorporates image segmentation as pre-processing for CNNs. We confirmed that the proposed patch-based pre-processing and post-processing were superior than Baseline 1.
For a detailed analysis, we performed Welch’s t-test [35] to the F-measures of ours and Baseline 1. As a result, we calculated a p-value of 0.000688, which statistically confirms the effectiveness of ours. Therefore, we can conclude that ours can accurately and robustly classify the river state. Fig 5 shows examples of classification results from ours with WideResNet. This figure shows that the appearance, such as color, varies between the images, according to the level of flood risk. Ours can suitably capture these differences in appearance, and perform successful classification.
Two top rows: images when a flood risk exists; two bottom rows: images in the normal case (no flood risk).
Discussion
We further discuss the effectiveness of our method in this section. From Table 3, it can be observed that our method increases the performance especially for WideResNet. This may be because this CNN model is suitable for small images. In fact, the original paper [19] shows that WideResNet is especially suitable for CIFAR10 [36], CIFAR100 [36] and street view house numbers (SVHN) [37], which include small images among many datasets. Also, for Baselines 1 and 2, Table 3 shows that the performance by ResNet101 [18] and ResNeXt [20] is low. This may be because these CNN models with many layers need much training data. Because ours can increase training data via the patch-based processing, we can successfully perform the river state classification even when using such CNN models.
Fig 6 shows examples that can be classified correctly by ours, but not by Baseline 1. Fig 6(b) and 6(c) show image of the same river. However, Fig 6(b) is a positive sample (flood risk case), whereas Fig 6(c) is a negative sample (normal case). Because the difference in appearance is small, these images may be difficult to classify correctly. However, ours can classify these images correctly because we can accurately train the river features by patch-based data augmentation in the training phase. For Fig 6(a), 6(b) and 6(d), Baseline 1 incorrectly detected some regions, such as the road and grassland as the river. This misclassification may occur because Baseline 1 trains and tests CNN models by using the incorrectly detected regions. In contrast, because ours performs ensemble learning —that is, patch selection in the training phase and majority voting in the test phase— accurate classification becomes feasible even if erroneously-detected water regions remain in the target image. As a result, we conclude that the adjustment of general CNNs to the river state classification becomes feasible by the proposed method.
Upper row: original images; lower row: segmentation results.
Conclusion
This paper proposed a method for classifying the river state (a flood risk exists or not) from river surveillance camera images by combining patch-based processing and CNN. Although CNN needs much training data, the number of river surveillance camera images is limited because flood does not frequently occur. Also, river surveillance camera images include objects that are irrelevant to the flood risk. Thus, the direct use of CNN may not work well for the river state classification. To overcome this limitation, we developed patch-based processing for adjusting CNN to the river state classification. By increasing training data and selecting relevant patches via the patch-based processing, the adjustment of general CNNs to the river state classification became feasible. We can develop the proposed patch-based processing and CNN independently. Therefore, we have the practical merits that any CNN can be used according to each user’s purposes, and the maintenance and improvement of each component of the whole system can be easily performed.
In the future, we will improve the ensemble algorithm by introducing boosting schemes such as [38, 39]. We are also interested in developing a new method for flood disaster prevention that can collaboratively use Web data such as weather information and Twitter posts as well as river surveillance camera images.
Acknowledgments
We thank Edanz Group (https://en-author-services.edanzgroup.com/) for editing a draft of this manuscript.
References
- 1. Um JH. Inquiry of special quality of supersonic sensor for water level by non-contact. The Bulletin of Korea Environment Engineers Association. 2000;162:30–36.
- 2. Yorozuya A, Tsukahara M, Kudo S, Koseki H, Fueta T. Development of water-level/velocity measuring instrument with radar technique. The Journal of Japan Society of Civil Engineers. 2016;72(5):305–311.
- 3.
Burt A, Lehmkuhl M, Burt CM, Styles SW. Water level sensor and datalogger testing and demonstration. Irrigation Training and Research Center, California Polytechnic State University. 2013;Report No. 99-002.
- 4.
Iwahashi M, Udomsiri S, Imai Y, Tansuriyavong S. Functionally layered coding for river monitoring. In: Proc. Int. Workshop on Advanced Image Technology (IWAIT); 2007. p. 918–923.
- 5. Minella JP, Merten GH, Reichert JM, Clarke RT. Estimating suspended sediment concentrations from turbidity measurements and the calibration problem. Hydrological Processes. 2008;22(12):1819–1830.
- 6.
Sarker C, Alvarez LM, Woodley A. Integrating recursive Bayesian estimation with support vector machine to map probability of flooding from multispectral Landsat data. In: Proc. Int. Conf. on Digital Image Computing: Techniques and Applications; 2016. p. 1–8.
- 7. Sarker C, Mejias L, Maire F, Woodley A. Integrating recursive Bayesian estimation with support vector machine to map probability of flooding from multispectral Landsat data. Remote Sensing. 2019;11(19):1–8.
- 8.
Takagi Y, Yoneoka T, Mori H, Yoda M, Tsujikawa A, Saito T. Development of a water level measuring system using image processing. In: Proc. IWA Conf. Instrumentation, Control and Automation; 2001. p. 309–316.
- 9.
Iwahashi M, Udomsiri S. Water level detection from video with FIR filtering. In: Proc. Int. Conf. Comput. Commun. and Networks; 2006. p. 826–831.
- 10.
Saito A, Iwahashi M. Water level detection algorithm based on synchronous frame addition and filtering. In: Proc. IEICE Circuits and Systems “Karuizawa” Workshop; 2006. p. 525–530.
- 11. Kim J, Han Y, Hahn H. Embedded implementation of image-based water-level measurement system. ITE Computer Vision. 2011;5(2):125–133.
- 12. Tsubaki R, Fujita I, Mama S, Takemura H, Kinbara K. Optimal viewing angle and image analysis method for measuring flood discharge of small-sized river using existing CCTV system. Advances in River Engineering. 2009;15:501–506.
- 13.
Oga T, Umeki Y, Iwahashi M, Matsuda Y. River water quality estimation based on convolutional neural network. In: Proc. Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC); 2018. p. 1305–1308.
- 14.
Krizhevsky A, Sutskever I, Hinton G. Imagenet classification with deep convolutional neural networks. In: Proc. Adv. in Neural Inf. Process. Syst.; 2012. p. 1097–1105.
- 15.
Lin M, Chen Q, Yan S. Network in network. arXiv preprint arXiv:13124400;2013.
- 16.
Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov D, et al. Going deeper with convolutions. In: Proc. IEEE Conf. Comput. Vis. Patt. Recognit.; 2015. p. 1–9.
- 17.
Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:14091556;2014.
- 18.
He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proc. IEEE Conf. Comput. Vis. Patt. Recognit.; 2016. p. 770–778.
- 19.
Zagoruyko S, Komodakis N. Wide Residual Networks. arXiv preprint arXiv:160507146;2016.
- 20.
Xie S, Girshick R, Dollar P, Tu Z, He K. Aggregated residual transformations for deep neural networks. In: Proc. IEEE Conf. Comput. Vis. Patt. Recognit.; 2017. p. 5987–5995.
- 21.
Zhao H, Shi J, Qi X, Wang X, Jia J. Pyramid scene parsing network. In: Proc. IEEE Conf. Comput. Vis. Patt. Recognit.; 2017. p. 2881–2890.
- 22.
Hou L, Samaras D, Kurc TM, Gao Y, Davis JE, Saltz JH. Patch-based convolutional neural network for whole slide tissue image classification. In: Proc. IEEE Conf. Comput. Vis. Patt. Recognit.; 2016. p. 2424–2433.
- 23. Natsev A, Rastogi R, Shim K. WALRUS: A similarity retrieval algorithm for image databeses. IEEE Trans Knowledge and Data Engineering. 2004;16(3):301–316.
- 24. Gomez L, Nicolau A, Karatzas D. Improving patch-based scene text script identification with ensembles of conjoined networks. Pattern Recognition. 2017;67:85–96.
- 25. Zhou B, Zhao H, Puig X, Xiao T, Fidler S, Barriuso A. Semantic understanding of scenes through the ADE20K dataset. International Journal of Computer Vision. 2019;127(3):302–321.
- 26.
Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation. In: Proc. IEEE Conf. Comput. Vis. Patt. Recognit.; 2015. p. 3431–3440.
- 27. Badrinarayanan V, Kendall A, Cipolla R. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans Pattern Analysis and Machine Intelligence. 2017;39(12):2481–2495. pmid:28060704
- 28. Zhou B, Lapedriza A, Khosla A, Oliva A, Torralba A. Places: A 10 million image database for scene recognition. IEEE Trans Pattern Analysis and Machine Intelligence. 2017;40(6):1452–1464. pmid:28692961
- 29.
Adam C, Ng A, Lee H. An analysis of single-layer networks in unsupervised feature learning. In: Proc. Int. Conf. Artificial Intelligence and Statistics; 2011. p. 215–223.
- 30. Bengio Y, Simard P, Frasconi P. Learning long-term dependencies with gradient descent is difficult. IEEE Trans Neural Networks. 1994;5(2):157–166. pmid:18267787
- 31.
Bottou L. Stochastic gradient descent tricks. In: Neural networks: Tricks of the trade; 2012. p. 421–436.
- 32.
Anders K, John AH. A simple weight decay can improve generalization. In: Proc. Adv. in Neural Inf. Process. Syst. Morgan-Kaufmann; 1992. p. 950–957.
- 33.
He K, Zhang X, Ren S, Sun J. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: Proc. IEEE Int. Conf. Comput. Vis.; 2015. p. 1026–1034.
- 34.
Bishop CM. Pattern recognition and machine learning. Springer-Verlag; 2006.
- 35. Welch BL. The generalization of student’s problem when several different population variances are involved. Biometrika. 1947;34(1/2):28–35. pmid:20287819
- 36.
Krizhevsky A, Hinton G. Learning multiple layers of features from tiny images. Technical Report, Univ of Toronto;2009.
- 37.
Netzer Y, Wang T, Coates A, Bissacco A, Wu B, Ng AY. Reading digits in natural images with unsupervised feature learning. In: Proc. NIPS Workshop on Deep Learning and Unsupervised Feature Learning; 2011. p. 1–9.
- 38. Freund Y, Schapire RE. Improving patch-based scene text script identification with ensembles of conjoined networks. Journal of Computer and System Sciences. 1997;55(1):119–139.
- 39.
Walach E, Wolf L. Learning to count with CNN boosting. In: Proc. European Conf. Comput. Vis. (ECCV); 2016. p. 660–676.