深度學(xué)習(xí)中交叉熵

計(jì)算機(jī)視覺 (Computer Vision)

Deep Computer Vision is capable of doing object detection and image classification task. In image classification tasks, the particular system receives some input image and the system is aware of some predetermined set of categories or labels. There are some fixed set of category labels and the job of the computer is to look at the picture and assign it a fixed category label. Convolutional Neural Network (CNN) has gained wide popularity in the field of pattern recognition and machine learning. In our present work, we have constructed a Convolutional Neural Network (CNN) for the identification of the presence of tantalum and niobium fragments in a High Entropy Alloy (HEA). The results showed 100 % accuracy while testing the given dataset.

深度計(jì)算機(jī)視覺能夠執(zhí)行對象檢測和圖像分類任務(wù)。 在圖像分類任務(wù)中，特定系統(tǒng)接收一些輸入圖像，并且系統(tǒng)知道一些預(yù)定的類別或標(biāo)簽集。 有一些固定的類別標(biāo)簽集，計(jì)算機(jī)的工作是看圖片并為其分配一個(gè)固定的類別標(biāo)簽。 卷積神經(jīng)網(wǎng)絡(luò)(CNN)在模式識別和機(jī)器學(xué)習(xí)領(lǐng)域獲得了廣泛的普及。 在我們目前的工作中，我們構(gòu)建了卷積神經(jīng)網(wǎng)絡(luò)(CNN)，用于識別高熵合金(HEA)中鉭和鈮碎片的存在。 測試給定數(shù)據(jù)集時(shí)，結(jié)果顯示100％的準(zhǔn)確性。

Introduction

介紹

Vision is the most important sense that humans possess. In day to day life, people depend on vision for example identifying objects, picking objects, navigation, recognizing complex human emotions and behaviors. Deep computer vision is able to solve extraordinary complex tasks that were not able to be solved in the past. Facial detection and recognition and detection are an example of deep computer vision. Figure 1 shows the vision coming into a deep neural network in the form of images or pixels or videos and the output at the bottom is the depiction of a human face [1–4].

視覺是人類擁有的最重要的感覺。 在日常生活中，人們依靠視覺來識別物體，拾取物體，導(dǎo)航，識別復(fù)雜的人類情感和行為。 深入的計(jì)算機(jī)視覺能夠解決過去無法解決的非凡復(fù)雜任務(wù)。 面部檢測，識別和檢測是深度計(jì)算機(jī)視覺的一個(gè)示例。 圖1顯示了以圖像，像素或視頻形式進(jìn)入深層神經(jīng)網(wǎng)絡(luò)的視覺，底部的輸出是對人臉的描繪[1-4]。

Fig.1. Illustration of the working of Deep Computer Vision圖。1。 深度計(jì)算機(jī)視覺工作插圖

The next thing should be worth answering to the question, how computer process an image or a video, and how do they process pixels coming from those? The images are just numbers and also the pixels have some numerical values. So our image can be represented by a two-dimensional matrix consisting of numbers. Let’s understand this with an example of image identification i.e. whether the image is of a boy or a girl or an animal. Figure 2 shows that the output variable takes a class label and can produce a probability of belonging to a particular class.

接下來的事情應(yīng)該值得回答這個(gè)問題：計(jì)算機(jī)如何處理圖像或視頻，以及它們?nèi)绾翁幚韥碜赃@些圖像或視頻的像素？ 圖像只是數(shù)字，像素也有一些數(shù)值。 因此，我們的圖像可以由包含數(shù)字的二維矩陣表示。 讓我們以圖像識別的示例(即圖像是男孩還是女孩還是動物)來理解這一點(diǎn)。 圖2顯示了輸出變量帶有類別標(biāo)簽，并且可以產(chǎn)生屬于特定類別的概率。

Fig.2. Image Classification圖2。 影像分類

In order to properly classify the image, our pipeline must correctly tell about what is unique about the particular picture. Convolutional Neural Network (CNN) finds application in the manufacturing and material science domain. Lee et al. [5] proposed a CNN model for fault diagnosis and classification in the manufacturing process of semiconductors. Weimer et al. [6] designed deep convolutional neural network architectures for automated feature extraction in industrial applications. Scime et al. [7] used the CNN model for the detection of in situ processing defects in laser powder bed fusion additive manufacturing. The results showed that the CNN architecture improved the classification accuracy and overall flexibility of the designed system.

為了正確分類圖像，我們的管道必須正確告知特定圖片的獨(dú)特之處。 卷積神經(jīng)網(wǎng)絡(luò)(CNN)在制造和材料科學(xué)領(lǐng)域得到了應(yīng)用。 Lee等。 [5]提出了一種用于半導(dǎo)體制造過程中故障診斷和分類的CNN模型。 Weimer等。 [6]設(shè)計(jì)了用于工業(yè)應(yīng)用中自動特征提取的深度卷積神經(jīng)網(wǎng)絡(luò)體系結(jié)構(gòu)。 Scime等。 [7]使用CNN模型來檢測激光粉末床熔融增材制造中的原位加工缺陷。 結(jié)果表明，CNN體系結(jié)構(gòu)提高了設(shè)計(jì)系統(tǒng)的分類準(zhǔn)確性和整體靈活性。

In the present work, we have designed the CNN architecture for detecting the trace of tantalum and niobium in the microstructure of high entropy alloy (HEA). In 1995, Yeh et al. [8] firstly discovered the high entropy alloys, and in 2004 Cantor et al. [9] coined high entropy alloy as a multi-component system. HEAs are generally advanced alloys and novel alloys which are consist of 5–35 at.% where all the elements behave as principal elements. In comparison to their conventional alloys, they possess superior properties like high wear, corrosion resistance, high thermal stability, and high strength. Zhang et al. [10–11] listed down the various parameters for the parameters for fabrication of HEAs which are shown in the below equations:

在當(dāng)前的工作中，我們設(shè)計(jì)了CNN體系結(jié)構(gòu)，用于檢測高熵合金(HEA)微觀結(jié)構(gòu)中的痕量鉭和鈮。 1995年，Yeh等人。 [8]首先發(fā)現(xiàn)了高熵合金，2004年Cantor等人。 [9]創(chuàng)造了高熵合金作為多組分系統(tǒng)。 HEA通常是高級合金和新型合金，由5–35 at。％的成分組成，其中所有元素均作為主要元素。 與常規(guī)合金相比，它們具有優(yōu)異的性能，如高耐磨性，耐腐蝕性，高熱穩(wěn)定性和高強(qiáng)度。 張等。 [10-11]列出了制造HEA的各種參數(shù)，這些參數(shù)如下式所示：

HEAs find application in various industries like aerospace, submarines, automobiles, and nuclear power plant industries [12–14]. HEAs are also used as a filler material for the micro-joining process [15]. Geanta et al. [16] carried out the testing and characterization of HEAs from AlCrFeCoNi System for Military Applications. It was observed that at the melt state, the microstructure of HEAs has frozen appearance as shown in Figure 3.

HEA在航空航天，潛艇，汽車和核電廠等各種行業(yè)中都有應(yīng)用[12-14]。 HEA還用作微連接過程的填充材料[15]。 Geanta等。 [16]進(jìn)行了軍事應(yīng)用AlCrFeCoNi系統(tǒng)的HEA的測試和表征。 觀察到，在熔融狀態(tài)下，HEA的微觀結(jié)構(gòu)具有凍結(jié)外觀，如圖3所示。

Fig.3. The appearance of frozen microstructure圖3。 冷凍組織的外觀

Material and Methods

材料與方法

Geanta et al. [17] fabricated biocompatible FeTaNbTiZrMo HEAs. In our study, we have used microstructure data from their research. The obtained microstructure is shown in Figures 4 and 5. Data collection is the process of gathering and measuring information from countless different sources. In order to use the data we collect to develop practical artificial intelligence (AI) and machine learning solutions, it must be collected and stored in a way that makes sense for the business problem at hand. Since we had a shortage of images, so we first did Image Augmentation.

Geanta等。 [17]制作了生物相容的FeTaNbTiZrMo HEA。 在我們的研究中，我們使用了他們研究的微觀結(jié)構(gòu)數(shù)據(jù)。 獲得的微觀結(jié)構(gòu)如圖4和5所示。數(shù)據(jù)收集是從無數(shù)不同來源收集和測量信息的過程。 為了使用我們收集的數(shù)據(jù)來開發(fā)實(shí)用的人工智能(AI)和機(jī)器學(xué)習(xí)解決方案，必須以對眼前的業(yè)務(wù)問題有意義的方式來收集和存儲數(shù)據(jù)。 由于圖像不足，因此我們首先進(jìn)行圖像增強(qiáng)。

Fig.4.Undissolved Ta and Nb fragments in the FeTaNbTiZrMo alloy [17].圖4. FeTaNbTiZrMo合金中未溶解的Ta和Nb碎片[17]。 Fig.5.Undissolved tantalum fragment in the FeTaNbTiZrMo alloy.圖5 FeTaNbTiZrMo合金中未溶解的鉭碎片

Image data augmentation is used to expand the training dataset in order to improve the performance and ability of the model to generalize. Image data augmentation is supported in the Keras deep learning library via the Image Data Generator class. So, input data consists of two images. As we know that we can’t train our deep neural network with only two images because that would result in the over-fitting of the model. Over-fitting a model basically means that our model will give the best score on training data but not on testing or validation data or the data that it has not seen before. So such an over-fitted model will be of no use to train our model effectively, we will make more images with the help of these input images. We will achieve this by Image Augmentation.

圖像數(shù)據(jù)增強(qiáng)用于擴(kuò)展訓(xùn)練數(shù)據(jù)集，以提高模型的性能和泛化能力。 Keras深度學(xué)習(xí)庫通過圖像數(shù)據(jù)生成器類支持圖像數(shù)據(jù)增強(qiáng)。 因此，輸入數(shù)據(jù)包含兩個(gè)圖像。 眾所周知，我們無法僅使用兩個(gè)圖像來訓(xùn)練我們的深度神經(jīng)網(wǎng)絡(luò)，因?yàn)檫@將導(dǎo)致模型的過度擬合。 過度擬合模型基本上意味著我們的模型將在訓(xùn)練數(shù)據(jù)上給出最佳分?jǐn)?shù)，而在測試或驗(yàn)證數(shù)據(jù)或之前從未見過的數(shù)據(jù)上則給出最佳分?jǐn)?shù)。 因此，這種過度擬合的模型對于有效地訓(xùn)練我們的模型將毫無用處，我們將在這些輸入圖像的幫助下制作更多圖像。 我們將通過圖像增強(qiáng)來實(shí)現(xiàn)。

Fig.5.Undissolved tantalum fragment in the FeTaNbTiZrMo alloy.

圖5 FeTaNbTiZrMo合金中未溶解的鉭碎片

We can use the Image Data Generator class to achieve this. First, we will make the object of this class. After that we will provide some parameters that are basically the fluctuations or feature that we want to provide the image like luminous intensity, width shift range, height shift range, etc. and we can iterate over the directory where the images are kept in, by providing the path in the function. In this way, we can generate numerous data. In this project, we have generated approximately 3000 images for each image.

我們可以使用Image Data Generator類來實(shí)現(xiàn)這一點(diǎn)。 首先，我們將成為此類的對象。 之后，我們將提供一些基本參數(shù)，這些參數(shù)基本上是我們想要提供圖像的波動或特征，例如發(fā)光強(qiáng)度，寬度偏移范圍，高度偏移范圍等。我們可以通過以下方式遍歷保存圖像的目錄：在函數(shù)中提供路徑。 這樣，我們可以生成大量數(shù)據(jù)。 在此項(xiàng)目中，我們?yōu)槊總€(gè)圖像生成了大約3000張圖像。

We created two datasets for the training and testing purpose. Python programming was used for the development of the code required for constructing the Convolutional Neural Network architecture. A Convolutional Neural Network (ConvNet/CNN) is a Deep Learning algorithm that can take in an input image, assign importance (learnable weights and biases) to various aspects/objects in the image, and be able to differentiate one from the other. The pre-processing required in a ConvNet is much lower as compared to other classification algorithms. While in primitive methods filters are hand-engineered, with enough training, ConvNets have the ability to learn these filters/characteristics.

我們?yōu)橛?xùn)練和測試目的創(chuàng)建了兩個(gè)數(shù)據(jù)集。 Python編程用于開發(fā)構(gòu)建卷積神經(jīng)網(wǎng)絡(luò)體系結(jié)構(gòu)所需的代碼。 卷積神經(jīng)網(wǎng)絡(luò)(ConvNet / CNN)是一種深度學(xué)習(xí)算法，可以吸收輸入圖像，為圖像中的各個(gè)方面/對象分配重要性(可學(xué)習(xí)的權(quán)重和偏差)，并能夠區(qū)分彼此。 與其他分類算法相比，ConvNet中所需的預(yù)處理要低得多。 在原始方法中，過濾器是手工設(shè)計(jì)的，經(jīng)過足夠的培訓(xùn)，ConvNets可以學(xué)習(xí)這些過濾器/特征。

Results and Discussions

結(jié)果和討論

The augmented image of the microstructure is shown in Figure 6.

顯微組織的放大圖如圖6所示。

Fig.6.Augmented images of the microstructure圖6顯微組織的放大圖像

Model is compiled with loss-Binary cross-entropy and metrics-accuracy and optimizer is adam. To prevent the model from Over-fitting, early stopping and model checkpoints are used so as to prevent a model from overtraining. Early Stopping is basically a process in which the model is stopped training when it doesn’t undergo any improvement. This parameter is provided in early stopping while making its object. This parameter is known as Patience. Metrics and mode are also provided as a parameter to test the model on the basis of that. Suppose metrics are value accuracy and mode is maximum, so when the model will not show any improvement (increment in value accuracy), it will wait till the patience parameter and after that, it will stop. The results were quite satisfactory when we trained our model against unlabelled images. As we can see in Figure 7, during prediction, almost every actual value is matched with predicted value so our model has been trained effectively.

使用損失-二進(jìn)制交叉熵和度量準(zhǔn)確性來編譯模型，并且優(yōu)化器是亞當(dāng)。 為了防止模型過度擬合，使用了早期停止和模型檢查點(diǎn)，以防止模型過度訓(xùn)練。 基本上，“早期停止”是一個(gè)過程，其中模型在未進(jìn)行任何改進(jìn)時(shí)就停止訓(xùn)練。 在使其成為對象的早期停止中提供此參數(shù)。 此參數(shù)稱為耐心。 度量和模式也作為參數(shù)提供，以在此基礎(chǔ)上測試模型。 假設(shè)度量標(biāo)準(zhǔn)是值準(zhǔn)確性，并且模式是最大，那么當(dāng)模型沒有顯示出任何改善(值準(zhǔn)確性增加)時(shí)，它將等待直到耐心參數(shù)，然后才停止。 當(dāng)我們針對未標(biāo)記圖像訓(xùn)練模型時(shí)，結(jié)果非常令人滿意。 如圖7所示，在預(yù)測期間，幾乎每個(gè)實(shí)際值都與預(yù)測值匹配，因此我們的模型已得到有效訓(xùn)練。

Fig.7. The predicted value matches the Actual Value圖7。 預(yù)測值與實(shí)際值匹配

The graphs in Figure 8 show the changes in metrics while training. As we can see, the model loss is getting lower as the epoch increases and accuracy is increasing as the epoch increases.

圖8中的圖形顯示了訓(xùn)練期間指標(biāo)的變化。 如我們所見，隨著歷時(shí)的增加，模型損失越來越小，隨著歷時(shí)的增加，模型的準(zhǔn)確性也越來越高。

Fig.8.Graph showing model loss and model accuracy圖8顯示模型損失和模型準(zhǔn)確性的圖

Conclusion

結(jié)論

It can be concluded that the current research is basically about image processing and classification, in which we first collected data due to a shortage of data, we did data augmentation to train our deep learning model, after that, we implemented our model architecture and compilation is done. After training, the results are shown. It is observed that the predicted value matches the actual value resulting in good accuracy for the image classification of the fragments present in HEAs.

可以得出結(jié)論，當(dāng)前的研究基本上是關(guān)于圖像處理和分類的，其中我們首先由于數(shù)據(jù)不足而收集數(shù)據(jù)，我們進(jìn)行了數(shù)據(jù)擴(kuò)充以訓(xùn)練我們的深度學(xué)習(xí)模型，之后，我們實(shí)現(xiàn)了模型架構(gòu)和編譯已經(jīng)完成了。 訓(xùn)練后，將顯示結(jié)果。 可以看出，預(yù)測值與實(shí)際值匹配，從而導(dǎo)致HEA中存在的碎片的圖像分類具有良好的準(zhǔn)確性。

[1] Forsyth, David A., and Jean Ponce. Computer vision: a modern approach. Prentice Hall Professional Technical Reference, 2002.

[1] Forsyth，David A.和Jean Ponce。 計(jì)算機(jī)視覺：一種現(xiàn)代方法。 Prentice Hall專業(yè)技術(shù)參考，2002年。

[2] Mundy, J.L. and Zisserman, A. eds., 1992. Geometric invariance in computer vision (Vol. 92). Cambridge, MA: MIT press.

[2] Mundy，JL和Zisserman，A。編輯，1992。計(jì)算機(jī)視覺中的幾何不變性(第92卷)。 馬薩諸塞州劍橋市：麻省理工學(xué)院出版社。

[3] Bradski, G. and Kaehler, A., 2008. Learning OpenCV: Computer vision with the OpenCV library. “ O’Reilly Media, Inc.”.

[3] Bradski，G。和Kaehler，A.，2008年。學(xué)習(xí)OpenCV：使用OpenCV庫進(jìn)行計(jì)算機(jī)視覺。 “ O'Reilly Media，Inc.”。

[4] Schalkoff, R.J., 1989. Digital image processing and computer vision (Vol. 286). New York: Wiley.

[4] Schalkoff，RJ，1989。數(shù)字圖像處理和計(jì)算機(jī)視覺(第286卷)。 紐約：威利。

[5] Lee, K.B., Cheon, S. and Kim, C.O., 2017. A convolutional neural network for fault classification and diagnosis in semiconductor manufacturing processes. IEEE Transactions on Semiconductor Manufacturing, 30(2), pp.135–142.

[5] Lee，KB，Cheon，S.和Kim，CO，2017年。用于半導(dǎo)體制造過程中故障分類和診斷的卷積神經(jīng)網(wǎng)絡(luò)。 IEEE Transactions on Semiconductor Manufacturing，30(2)，第135–142頁。

[6] Weimer, D., Scholz-Reiter, B. and Shpitalni, M., 2016. Design of deep convolutional neural network architectures for automated feature extraction in industrial inspection. CIRP Annals, 65(1), pp.417–420.

[6] Weimer，D.，Scholz-Reiter，B.和Shpitalni，M.，2016年。用于工業(yè)檢測中自動特征提取的深度卷積神經(jīng)網(wǎng)絡(luò)架構(gòu)設(shè)計(jì)。 CIRP年鑒，65(1)，第417–420頁。

[7] Scime, L. and Beuth, J., 2018. A multi-scale convolutional neural network for autonomous anomaly detection and classification in a laser powder bed fusion additive manufacturing process. Additive Manufacturing, 24, pp.273–286.

[7] Scime，L.和Beuth，J.，2018年。一種多尺度卷積神經(jīng)網(wǎng)絡(luò)，用于在激光粉末床熔合添加劑制造過程中進(jìn)行自主異常檢測和分類。 增材制造，第24頁，第273-286頁。

[8] Yeh JW, Chen SK, Lin SJ, et al. Nanostructured high-entropy alloys with multiple principal elements: Novel alloy design concepts and outcomes. Advanced Engineering Materials. 2004;6(5):299–303

[8] Ye JW，Chen SK，Lin SJ等。 具有多個(gè)主要元素的納米結(jié)構(gòu)高熵合金：新穎的合金設(shè)計(jì)理念和成果。 先進(jìn)的工程材料。 2004; 6(5)：299–303

[9] Cantor B. High-entropy alloys. In: Buschow KHJ, Cahn RW, Flemings MC, Ilschner B, Kramer EJ, Mahajan S, et al. editors. Encyclopedia of Materials: Science and Technology. ISBN 978–0–08043152–9

[9] Cantor B.高熵合金。 在：Buschow KHJ，Cahn RW，Flemings MC，Ilschner B，Kramer EJ，Mahajan S等人中。 編輯。 材料百科全書：科學(xué)和技術(shù)。 ISBN 978–0–08043152–9

[10] Yeh JW, Chen YL, Lin SJ, et al. High-entropy alloys — A new era of exploitation. Materials Science Forum. 2007;560:1–9

[10] Ye JW，Chen YL，Lin SJ等。 高熵合金—開采的新紀(jì)元。 材料科學(xué)論壇。 2007; 560：1–9

[11] Zhang Y, Zhou YJ, Lin JP, et al. Solid-solution phase formation rules for multi-component alloys. Advanced Engineering Materials. 2008;10(6):534–538

[11]張勇，周永堅(jiān)，林建平，等。 多組分合金的固溶相形成規(guī)則。 先進(jìn)的工程材料。 2008; 10(6)：534–538

[12] Zhou YJ, Zhang Y, Wang YL, et al. Solid solution alloys of AlCoCrFeNiTix with excellent room temperature mechanical properties. Applied Physics Letters. 2007;90(18):1904

[12]周英杰，張穎，王永利，等。 AlCoCrFeNiTix的固溶合金具有出色的室溫機(jī)械性能。 應(yīng)用物理快報(bào)。 2007; 90(18)：1904

[13] Senkov ON, Wilks GB, Scott JM, Miracle DB. Mechanical properties of Nb25Mo25Ta25W25 and V20Nb20Mo20Ta20W20 refractory high entropy alloys. Intermetallics. 2011;19:698–706

[13] Senkov ON，Wilks GB，Scott JM和Miracle DB。 Nb25Mo25Ta25W25和V20Nb20Mo20Ta20W20難熔高熵合金的力學(xué)性能。 金屬間化合物。 2011; 19：698–706

[14] Lin CM, Tsai HL. Evolution of microstructure, hardness, and corrosion properties of high-entropy Al0.5CoCrFeNi alloy. Intermetallics. 2011;19(3):288–294

[14] Lin CM，Tsai HL。 高熵Al0.5CoCrFeNi合金的組織，硬度和腐蝕性能的演變。 金屬間化合物。 2011; 19(3)：288-294

[15] Ashutosh Sharma (April 6th 2020). High-Entropy Alloys for Micro- and Nanojoining Applications [Online First], IntechOpen, DOI: 10.5772/intechopen.91166. Available from: https://www.intechopen.com/online-first/high-entropy-alloys-for-micro-and-nanojoining-applications

[15] Ashutosh Sharma(2020年4月6日)。 用于微連接和納米連接的高熵合金[在線優(yōu)先]，IntechOpen，DOI：10.5772 / intechopen.91166。 可從以下網(wǎng)址獲得： https : //www.intechopen.com/online-first/high-entropy-alloys-for-micro-and-nanojoining-applications

[16] Victor Geanta and Ionelia Voiculescu (October 23rd 2019). Characterization and Testing of High-Entropy Alloys from AlCrFeCoNi System for Military Applications [Online First], IntechOpen, DOI: 10.5772/intechopen.88622.

[16] Victor Geanta和Ionelia Voiculescu(2019年10月23日)。 AlCrFeCoNi系統(tǒng)用于軍事應(yīng)用的高熵合金的表征和測試[在線優(yōu)先]，IntechOpen，DOI：10.5772 / intechopen.88622。

[17] . Victor Geanta, Ionelia Voiculescu, Petrica Vizureanu and Andrei Victor Sandu (September 21st 2019). High Entropy Alloys for Medical Applications [Online First], IntechOpen, DOI: 10.5772/intechopen.89318.

[17]。 Victor Geanta，Ionelia Voiculescu，Petrica Vizureanu和Andrei Victor Sandu(2019年9月21日)。 用于醫(yī)療應(yīng)用的高熵合金[在線優(yōu)先]，IntechOpen，DOI：10.5772 / intechopen.89318。

翻譯自: https://medium.com/towards-artificial-intelligence/deep-computer-vision-for-the-detection-of-tantalum-and-niobium-fragments-in-high-entropy-alloys-5d0c2d8c988a

深度學(xué)習(xí)中交叉熵

總結(jié)

以上是生活随笔為你收集整理的深度学习中交叉熵_深度计算机视觉，用于检测高熵合金中的钽和铌碎片的全部內(nèi)容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網(wǎng)站內(nèi)容還不錯(cuò)，歡迎將生活随笔推薦給好友。