- 論文名稱：《MobileNet Based Apple Leaf Diseases Identification》
- 論文作者： Chao X , Sun G , Zhao H , et al.
- 發表期刊：Mobile Networks and Applications, 2020(10).
- 論文總結：
- Research Gap:
基于MobileNet的蘋果葉病害識別 - Importance：
MobileNet對2類蘋果病理葉片的ACC為73.5%,識別速度為每張圖片0.22s - Limitations：
蘋果葉部病理分類類別較少
沒有對MobileNet進行改進
- 論文地址： https://www.researchgate.net/publication/344004543_MobileNet_Based_Apple_Leaf_Diseases_Identification
- 論文目錄
- Abstract
- 1.Introduction
- 2 Related work
- 2.1 Traditional apple leaf diseases inspection method
- 2.2 Development of traditional convolutional neural networks
- 2.3 Development of lightweight convolutional neural networks
- 3 Dataset construction for obtaining stable identification results
- 4 The MobileNet model for apple leaf diseases identification
- 5 The high precision models for apple leaf diseases identification
- 5.1 The ResNet152 model for apple leaf diseases identification
- 5.2 The InceptionV3 model for apple leaf diseases inspection
- 6 Experiments and results
- 6.1 Experimental setup
- 6.2 Dataset
- 6.3 Experiment results and analysis
- 7 Conclusions and future works
Abstract
| 原文 | 譯文 |
|---|---|
?? Alternaria leaf blotch, and rust are two common types of apple leaf diseases that severely affect apple yield. A timely andeffective detection of apple leaf diseases is crucial for ensuring the healthy development of the apple industry. In general,these diseases are inspected by experienced experts one by one. This is a time-consuming task with unstable precision.Therefore, in this paper, we proposed a LOW-COST, STABLE, HIGH precision apple leaf diseases identification method.This is achieved by employing MobileNet model. Firstly, comparing with general deep learning model, it is a LOW-COST model because it can be easily deployed on mobile devices. Secondly, instead of experienced experts, everyone can finish the apple leaf diseases inspection STABLELY by the help of our algorithm. Thirdly, the precision of MobileNet is nearlythe same with existing complicated deep learning models. Finally, in order to demonstrated the effectiveness of ourproposed method, several experiments have been carried out for apple leaf diseases identification. We have compared the efficiency and precision with the famous CNN models: i.e. ResNet152 and InceptionV3. Here, the apple disease datasets (including classes: Alternaria leaf blotch and rust leaf) were collected by the agriculture experts in Shaanxi Province, China. | ?? 鏈格孢葉斑病和銹病是嚴重影響蘋果產量的兩種常見的蘋果葉病，及時有效地發現蘋果葉片病害，是保障蘋果產業健康發展的關鍵，一般來說，這些疾病都是由有經驗的專家一一檢查，這是一項耗時且精度不穩定的任務，因此，在本文中，我們提出了一種低成本、穩定、高精度的蘋果葉病識別方法，這是通過采用 MobileNet 模型實作的，首先，與一般的深度學習模型相比，它是一個低成本的模型，因為它可以很容易地部署在移動設備上，其次，在我們的演算法的幫助下，每個人都可以穩定地完成蘋果葉病檢查，而不是有經驗的專家，第三，MobileNet 的精度與現有復雜的深度學習模型幾乎相同，最后，為了證明我們提出的方法的有效性，對蘋果葉片病害的識別進行了多次實驗，我們已經將效率和精度與著名的 CNN 模型進行了比較：即 ResNet152 和 InceptionV3，這里，蘋果病害資料集（包括類：鏈格孢葉斑病和銹葉）由中國陜西省的農業專家收集， |
| Keywords: Keywords: Apple leaf diseases . Mobile device . MobileNet . Deep learning | 關鍵詞： 蘋果葉病；移動設備 ；MobileNet；深度學習 |
1.Introduction
| 原文 | 譯文 |
|---|---|
?? With a high nutritional and medicinal value, apple is one of the most productive types of fruit in the world. However,various diseases occur frequently on a large scale in apple production, such as Apple Alternaria leaf blotch (caused by Alternaria alternata f.sp. mali), and Apple rust (caused by Pucciniaceae glue rust), which affect the quality of fruits and thereby causing substantial economic losses. | ?? 蘋果具有很高的營養和藥用價值，是世界上產量最高的水果之一， 然而，蘋果生產中經常發生各種病害，如蘋果鏈格孢葉斑病（由馬里鏈格孢引起）、蘋果銹病（由柄銹菌引起），影響果實品質和品質， 從而造成重大經濟損失， |
| ?? Currently, the apple leaf diseases are mainly inspected by experienced experts. They need to check the apple leaves one by one. This is a huge job. The number of leaves for one apple tree is large enough. For a whole apple yield, we do not have enough experienced experts to finish such kind of inspection task. Furthermore, a large number of errors will be appeared when these experts become tired, especially for some similar leaf diseases. | ?? 目前，蘋果葉片病害主要由經驗豐富的專家進行檢查， 他們需要一張一張地檢查蘋果的葉子， 這是一項巨大的作業， 一棵蘋果樹的葉子數量足夠大， 對于整個蘋果的產量，我們沒有足夠的經驗豐富的專家來完成這樣的檢查任務， 再者，當這些專家累了，就會出現大量的錯誤，特別是對于一些類似的葉病， |
| ?? Therefore, we need an algorithm to help farmers to resolve this problem. This algorithm can let non-experience-farmers to identify these apple leaf diseases without the helps from experts. This algorithm should have the following 3 merits: ?? – Stable: the method should have the ability to identify the similar apple leaf diseases. ?? – Low Cost: the method could be simply used in mobile devices. ?? – Efficiency & Precision: the disease should be identified in less than one second with high precision. | ?? 因此，我們需要一種演算法來幫助農民解決這個問題， 該演算法可以讓沒有經驗的農民在沒有專家幫助的情況下識別這些蘋果葉病， 該演算法應具有以下3個優點： ?? – 穩定：該方法應具有識別相似蘋果葉病害的能力， ?? – 低成本：該方法可以簡單地用于移動設備， ?? – 效率和精度：應在不到一秒的時間內以高精度識別疾病， |
| ?? However, current existing researches cannot fully satisfy these three issues for apple leaf diseases identification. We divided existing approaches into two classifications: inspection by experienced experts, and inspection by using deep learning methods. Here, the withdraws of experienced experts’ inspection have been described in previous paragraph. On the other hand, deep learning method cannot be used for apple leaf diseases identification directly. This is because these models are too complex to be used on mobile device directly. | ?? 然而，目前現有的研究還不能完全滿足蘋果葉片病害識別的這三個問題， 我們將現有方法分為兩類：經驗豐富的專家檢查和使用深度學習方法的檢查， 這里，有經驗的專家檢查的退出在前一段已經描述過， 另一方面，深度學習方法不能直接用于蘋果葉片病害識別， 這是因為這些模型太復雜，無法直接在移動設備上使用， |
| ?? In this paper, we built a mobile-based model for apple leaf diseases identification based on MobileNet. This model is a mobile version CNN model. Its precision is nearly the same as the general CNN model. Meanwhile, its efficiency is high enough for apple leaf diseases identification. In our method, firstly, in order to let the disease identification as STABLE as possible, we have invited the agriculture experts from Chinese Academy of Agricultural Sciences, China for obtaining all kinds of leaves with different apple leaf diseases. These datasets are taken from Shaanxi Province, China. Note that, we have mainly used two kinds of diseases (Alternaria leaf blotch and Apple rust) to demonstrate the effectiveness of our proposed method. Secondly, in order to satisfy the LOWCOST issue, the MobileNet model is employed for apple leaf diseases identification. This is because it can be easily deployed on mobile devices. Finally, in order to achieve the goal of HIGH efficiency and precision, the MobileNet model is optimized according to the features of apple leaf diseases. Furthermore, we have also tried several other models. As a result, the MobileNet model is the best choice. A balance of efficiency and precision is achieved by using MobileNet for apple leaf diseases identification. | ?? 在本文中，我們基于 MobileNet 構建了一個基于移動的蘋果葉病識別模型，這個模型是手機版的CNN模型，它的精度與一般的CNN模型幾乎相同，同時，其效率對于蘋果葉片病害的識別也足夠高，在我們的方法中，首先，為了讓病害鑒定盡可能穩定，我們邀請了中國農業科學院的農業專家來獲取各種不同蘋果葉片病害的葉子，這些資料集來自中國陜西省，請注意，我們主要使用兩種疾病（鏈格孢葉斑病和蘋果銹病）來證明我們提出的方法的有效性，其次，為了滿足低成本問題，采用MobileNet模型進行蘋果葉片病害識別，這是因為它可以輕松部署在移動設備上，最后，為了達到高效、精準的目標，根據蘋果葉片病害的特點對MobileNet模型進行了優化，此外，我們還嘗試了其他幾種模型，因此，MobileNet 模型是最佳選擇，通過使用 MobileNet 進行蘋果葉病害識別，實作了效率和精度的平衡， |
| ?? The remainder of this paper is organized as follows. Section 2 provides a brief survey on apple leaf diseases identification. Section 3 introduces the dataset construction for obtaining stable identification results. Section 4 describes MobileNet based apple leaf diseases identification method; followed by another two famous CNN models: i.e. ResNet152 and InceptionV3 for comparing with MobileNet model on apple leaf diseases inspection in Section 5. Section 6 discusses the effectiveness of our method through comparing with ResNet152 and InceptionV3 models. Section 7 concludes the paper with future extensions. | ?? The remainder of this paper is organized as follows. Section 2 provides a brief survey on apple leaf diseases identification. Section 3 introduces the dataset construction for obtaining stable identification results. Section 4 describes MobileNet based apple leaf diseases identification method; followed by another two famous CNN models: i.e. ResNet152 and InceptionV3 for comparing with MobileNet model on apple leaf diseases inspection in Section 5. Section 6 discusses the effectiveness of our method through comparing with ResNet152 and InceptionV3 models. Section 7 concludes the paper with future extensions. |
2 Related work
| 原文 | 譯文 |
|---|---|
| ?? In this section, we will simply introduce the traditional apple leaf diseases identification method in Section 2.1. The development of traditional Convolutional Neural Networks is described in Section 2.2, followed by the lightweight Convolutional Neural Networks in Section 2.3. In this paper, we need a balance of precision and efficiency. Therefore, a lightweight model (MobileNet) is employed. Because its precision is nearly the same with that of traditional CNN, while its efficiency is very high for deploying on mobile devices. | ?? 在本節中，我們將在2.1節中簡單介紹傳統的蘋果葉片病害識別方法， 傳統卷積神經網路的發展在 2.2 節中描述，然后是輕量級卷積神經網路在 2.3 節中， 在本文中，我們需要在精度和效率之間取得平衡， 因此，采用了輕量級模型（MobileNet）， 因為它的精度與傳統的CNN幾乎相同，而在移動設備上部署的效率非常高， |
2.1 Traditional apple leaf diseases inspection method
| 原文 | 譯文 |
|---|---|
?? Traditionally, the visual observation by experienced experts has been carried out to diagnose plant diseases. However, recognition efficiency and accuracy are extremely unstable [1], due to subjective perception. Actually, such kind of traditional method is difficult to be completely implemented. This is because it is a huge job, even for inspecting the leaves of one apple tree. Meanwhile, such kind of job can only be finished by experienced experts. Furthermore, the inspection process is unstable because the experts will become tired. Therefore, the traditional apple leaf diseases inspection method should be replaced by some STABLE, LOW-COST, HIGH-EFFICIENCY-PRECISION methods. Deep learning can be used for this task since it has been well developed in recent years, which will be introduced in the next section. | ?? 傳統上，由經驗豐富的專家進行目視觀察來診斷植物病害， 然而，由于主觀感知，識別效率和準確性極不穩定[1]， 實際上，這種傳統的方法很難完全實作， 這是因為這是一項艱巨的作業，即使是檢查一棵蘋果樹的葉子也是如此， 同時，這種作業只能由經驗豐富的專家來完成， 此外，檢查程序不穩定，因為專家會變得疲倦， 因此，傳統的蘋果葉病檢測方法應該被一些穩定、低成本、高效率、精確的方法所取代， 深度學習可以用于這項任務，因為它近年來得到了很好的發展，這將在下一節中介紹， |
2.2 Development of traditional convolutional neural networks
| 原文 | 譯文 |
|---|---|
| ?? Convolutional neural network (CNN) [2] is the most commonly method applied to image-based recognition tasks in the deep learning framework. It uses relatively little preprocessing compared to other image classification algorithms and automatically learns the internal features of massive images. | ?? 卷積神經網路 (CNN) [2] 是深度學習框架中最常應用于基于影像的識別任務的方法， 與其他影像分類演算法相比，它使用相對較少的預處理，并自動學習海量影像的內部特征， |
?? At the beginning, limited to the processing ability of the hardware, CNNs cannot achieve high efficiency and precision. Such as Vishnupriya et al. [3] used convolution neural network to train a music dataset which includes ten different genres. The proposed system classified music into various genres by extracting the feature vector which was Mel Frequency Cepstral Coefficient (MFCC). The results showed that the accuracy level was around 76%. Lee et al. [4] developed a method to classify DNA damage patterns on comet assay images including detection, adjustment and analysis, reaching an average classification accuracy of 86.80% for 20 test datasets including more than 300 images. | ?? 最初，受限于硬體的處理能力，CNNs 無法實作高效率和高精度， 如 Vishnupriya 等人 [3] 使用卷積神經網路來訓練包含十種不同流派的音樂資料集， 所提出的系統通過提取作為梅爾頻率倒譜系數（MFCC）的特征向量將音樂分類為各種流派， 結果表明，準確率約為 76%， 李等人 [4] 開發了一種對彗星分析影像上的 DNA 損傷模式進行分類的方法，包括檢測、調整和分析，對 20 個測驗資料集（包括 300 多幅影像）的平均分類準確率達到 86.80%， |
| ?? Then, with the development of hardware, CNNs can be trained on large datasets, thus effectively improved the accuracy of the model. For example, Jiang et al. [5] proposed a new real-time apple leaf diseases detection model that is based on deep-CNNs. In the experiments, on a dataset of 26,377 images of diseases apple leaves (including classes: alternaria leaf blotch leaf, brow spot leaf, mosaic leaf, grey spot leaf and rust leaf), this approach realized a detection performance of 78.80% mAP. Verma et al. [6] employed CNN to develop a facial emotion recognition model to categorizes a facial expression into seven different emotions categorized. Matsubara et al. [7] proposed a CNN approach that combined spectral clustering information processing to classify lung cancer. Jaiswal et al. [8] built a model for classification of sound based on CNN, this model can be used for deforestation detection, Gunshot detection in urban areas and also for detecting unusual sounds in streets like a cry for help, tyres screeching etc. at odd hours. Francis et al. [9] created and developed a convolution neural network model for the detection and classification of plant diseases from apple and tomato leaf images, which trained on a dataset containing 3663 images and attained an accuracy of 87%. | ?? 然后，隨著硬體的發展，CNNs可以在大資料集上進行訓練，從而有效地提高了模型的準確性，例如，蔣等人[5] 提出了一種新的基于 deep-CNN 的實時蘋果葉病檢測模型，實驗中，在26,377張病害蘋果葉片影像資料集（包括：鏈格孢葉斑葉、斑葉、花葉、灰斑葉和銹葉）的資料集上，該方法實作了78.80%的mAP檢測性能，維爾馬等人 [6] 使用 CNN 開發面部情緒識別模型，將面部表情分為七種不同的情緒分類，松原等人 [7]提出了一種結合光譜聚類資訊處理的CNN方法對肺癌進行分類，賈斯瓦爾等 人[8] 基于 CNN 建立了一個聲音分類模型，該模型可用于森林砍伐檢測、城市地區的槍聲檢測，也可用于檢測街道上不尋常的聲音，如求救聲、輪胎尖叫聲等弗朗西斯等人 [9] 創建并開發了一個卷積神經網路模型，用于從蘋果和番茄葉影像中檢測和分類植物病害，該模型在包含 3663 個影像的資料集上進行訓練，準確率達到 87%， |
| ?? In order to further improve the accuracy and efficiency of the CNNs, the optimization of CNNs has become a hot topic in recent years. Xiong et al. [10] proposed a cross-connected CNN for traffic sign recognition. The experimental results on well-known dataset showed that the algorithm improved the accuracy. Sajjad et al. [11] built a CNN for facial expression recognition, with a final accuracy of 93.39%. Hu et al. [12] presented an integrated optimization method of simulated annealing (SA) and Gaussian convolution based on traditional Convolutional Neural Network. In the experiment, this method was applied to the MNIST and CIFAR-10 databased. Agarwal et al. [13] developed a convolution neural network based approach to identify the disease in apple fruit, which achieved the state of the art accuracy of 99%. Nachtigall et al. [14] studied the use of convolutional neural networks to automatically detect and classify diseases, nutritional deficiencies and damage by herbicides on apple trees from a dataset consisting of 2359 images of their leaves, achieving a 97.3% accuracy. | ?? 為了進一步提高CNNs的準確性和效率，CNNs的優化成為近年來的熱門話題，熊等人 [10] 提出了一種用于交通標志識別的交叉連接 CNN，在知名資料集上的實驗結果表明，該演算法提高了準確率，薩賈德等人 [11]構建了一個用于面部表情識別的CNN，最終準確率為93.39%，胡等人 [12]提出了一種基于傳統卷積神經網路的模擬退火（SA）和高斯卷積的綜合優化方法，在實驗中，該方法應用于MNIST和CIFAR-10資料庫，阿加瓦爾等人 [13] 開發了一種基于卷積神經網路的方法來識別蘋果果實中的疾病，達到了 99% 的最新準確率， Nachtigall 等人 [14] 研究了使用卷積神經網路從包含 2359 張葉子影像的資料集中自動檢測和分類蘋果樹上的疾病、營養缺乏和除草劑造成的損害，準確率達到 97.3%， |
| ?? In order to achieve higher accuracy, the general development trend of CNNs has been to make deeper and more complicated networks. However, the complex structure of models cannot make the network more efficient in terms of speed and size. | ??為了達到更高的準確率，CNNs的總體發展趨勢是制造更深、更復雜的網路， 然而，模型的復雜結構并不能使網路在速度和規模方面更加高效， |
2.3 Development of lightweight convolutional neural networks
| 原文 | 譯文 |
|---|---|
| ?? With the rapid spread of smart portable devices and the development of mobile service technology, the possibility of introducing smart applications in mobile environments is receiving increased attentions [15]. Due to the limitations of hardware and computing ability of mobile devices, object detection and recognition based on traditional CNNs is difficult to complete in ideal times, for example human detection in a video [16,17], self-driving car [18–20], robotics and automation [21],evacuation [22, 23], etc. | ?? 隨著智能便攜設備的迅速普及和移動服務技術的發展，在移動環境中引入智能應用的可能性越來越受到關注[15]， 由于移動設備硬體和計算能力的限制，基于傳統CNNs的物體檢測和識別難以在理想時間完成，例如視頻中的人體檢測[16,17]、自動駕駛汽車[18-20] ]、機器人與自動化 [21]、疏散 [22、23] 等， |
?? In order to combine the power of CNN with specific production practice, and make it more practical and usable, many lightweight Convolutional Neural Networks were proposed. Lo et al. [24] applied a deep CNN with Xception model to perform malware image classification, with a validation accuracy as high as 99.17%. TR et al. [25] proposed a hybrid Xception model for human protein atlas image classification, the F1 Score obtained by this model was 0.69. Nakamichi et al. [26] classified Circulating Tumor Cells (CTC) by SqueezeNet. A dataset consisting of 5040 microscopy images (6 samples) was used to evaluate the effectiveness. The experimental results demonstrated that the accuracy of the proposed method up to 89.86%. Hidayatuloh et al. [27] built a model based on squeezenet architecture to classify seven types of tomato plant diseases on the leaves including healthy leaves, with an average accuracy of identification of 86.92%. The MobileNet model we use in this work has also received a lot of attention in recent years. Rabano et al. [28] built a model based on MobileNet [29] for common trash classification. A dataset consisting of 2527 trash images was used for the training, with a final accuracy of 87.2%. Gavai et al. [30] applied MobileNet to flower classification. The experimental performance showed that the model based on MobileNet can greatly minimize the time and space for flower classification compromising the accuracy slightly. Xu et al. [31] modified and reduced the MobileNet structures to train the datasets from IEEE Detection and Classification of Acoustic Scenes and Events (DCASE). With this model, they succeed in reaching the validation rate of 75.99%. Sae-Lim et al. [32] proposed a modified MobileNet for skin lesion classification. The official dataset of Human Against Machine with 10,000 training images was used to train this model. The results showed that the modified model had achieved a good performance. Liu et al. [33] proposed a method combining an embedded system and deep learning based on MobileNetV2 and transfer learning. The results showed that the MobilenetV2 model plus transfer learning could be a better choice for the real-time classification of the marine animal images than InceptionV3 and MobilenetV1 models. | ?? 為了將CNN的強大功能與具體的生產實踐相結合，使其更加實用和好用，許多輕量級的卷積神經網路被提出，洛等人 [24] 應用帶有 Xception 模型的深度 CNN 來執行惡意軟體影像分類，驗證準確率高達 99.17%， TR 等 [25]提出了一種用于人類蛋白質圖譜影像分類的混合Xception模型，該模型得到的F1 Score為0.69，中道等人 [26] 通過 SqueezeNet 對回圈腫瘤細胞 (CTC) 進行分類，使用由 5040 張顯微鏡影像（6 個樣本）組成的資料集來評估有效性，實驗結果表明，該方法的準確率高達89.86%， Hidayatuloh 等 [27]基于squeezenet架構建立了一個模型，對包括健康葉片在內的7種番茄植物病害進行分類，平均識別準確率為86.92%，我們在這項作業中使用的 MobileNet 模型近年來也受到了很多關注，拉巴諾等人 [28] 建立了一個基于 MobileNet [29] 的模型，用于常見垃圾分類，訓練使用由 2527 張垃圾影像組成的資料集，最終準確率為 87.2%，加瓦伊等人 [30] 將 MobileNet 應用于花卉分類，實驗性能表明，基于 MobileNet 的模型可以極大地減少花卉分類的時間和空間，但對準確性略有影響，徐等人 [31] 修改和減少了 MobileNet 結構，以訓練來自 IEEE 聲學場景和事件檢測和分類 (DCASE) 的資料集，使用這個模型，他們成功地達到了 75.99% 的驗證率， Sae-Lim 等人 [32] 提出了一種改進的 MobileNet 用于皮膚病變分類，使用具有 10,000 張訓練影像的 Human Against Machine 官方資料集來訓練該模型，結果表明，改進后的模型取得了良好的性能，劉等人 [33] 提出了一種基于 MobileNetV2 和遷移學習的將嵌入式系統和深度學習相結合的方法，結果表明，與 InceptionV3 和 MobilenetV1 模型相比，MobilenetV2 模型加上遷移學習可以成為海洋動物影像實時分類的更好選擇， |
| ?? In this paper, we have tried three kinds of models: ResNet152, InceptionV3 and MobileNet. Here, ResNet152 belongs to traditional CNNs as introduced above. The other two lightweight models have nearly the same precision with ResNet152, while their efficiency is much higher than ResNet152. Therefore, through comparing the identification results of these three models, we choose MobileNet as our apple leaf diseases identification model. The details will be introduced in the following two Sections. | ?? 在本文中，我們嘗試了三種模型：ResNet152、InceptionV3 和 MobileNet， 這里，ResNet152 屬于上面介紹的傳統 CNN， 另外兩個輕量級模型的精度與 ResNet152 幾乎相同，而它們的效率遠高于 ResNet152， 因此，通過比較這三個模型的識別結果，我們選擇MobileNet作為我們的蘋果葉病識別模型， 詳細內容將在以下兩節介紹， |
3 Dataset construction for obtaining stable identification results
| 原文 | 譯文 |
|---|---|
| ?? Currently, apple leaf diseases are mainly inspected by experienced experts. They need to check apple leaves one by one. However, this is an impossible task. Because even the number of apple leaves in one tree is too large to be checked by person. In this case, experts can only check a very small part of apple trees in an apple yield. The apple leaf diseases can only be predicted by experts’ experiences according to the small number of sample checking. This will lead to a very unstable inspection result. | ?? 目前，蘋果葉片病害主要由有經驗的專家進行檢查， 他們需要一張一張地檢查蘋果的葉子， 然而，這是一項不可能完成的任務， 因為即使是一棵樹上的蘋果葉子數量也太大，無法由人來檢查， 在這種情況下，專家只能檢查蘋果產量中的一小部分蘋果樹， 蘋果葉片病害只能靠專家經驗，根據樣本檢查數量少來預測， 這將導致非常不穩定的檢查結果， |
| ?? On the other hand, the apple leaf diseases identification can only be carried out by very few experts. Their burden is too large. When they become tired, wrong apple leaf diseases will be judged, especially for the apple leaf with very similar image but totally different apple leaf diseases. Figure 1 is such an example. There are two kinds of apple leaf diseases: apple Alternaria leaf blotch and rust. Here, the first image and the last image are rust; while the remainder four images are apple Alternaria leaf blotch. However, the second image is easily to be judged as rust; the third image cannot be easily to be judged as apple Alternaria leaf blotch. | ?? 另一方面，蘋果葉片病害鑒定只能由極少數專家進行， 他們的負擔太大了， 當他們累了，就會判斷出錯誤的蘋果葉病，特別是對于影像非常相似但蘋果葉病完全不同的蘋果葉， 圖 1 就是這樣一個例子， 蘋果葉病有兩種：蘋果鏈格孢葉斑病和銹病， 在這里，第一個影像和最后一個影像是銹跡； 而其余四張影像是蘋果鏈格孢葉斑病， 然而，第二幅影像很容易被判斷為生銹； 第三張圖片不容易被判斷為蘋果鏈格孢葉斑病， |
| 原文 | 譯文 |
|---|---|
| ?? Fig. 1 Two kinds of apple leaf diseases: apple Alternaria leaf blotch and rust. Here, the first image and the last image are rust; while the remainder four images are apple Alternaria leaf blotch. However, the second image is easily to be judged as rust; the third image cannot be easily to be judged as apple Alternaria leaf blotch . | ?? 圖1 兩種蘋果葉病：蘋果鏈格孢葉斑病和銹病， 在這里，第一個影像和最后一個影像是銹跡； 而其余四張影像是蘋果鏈格孢葉斑病， 然而，第二幅影像很容易被判斷為生銹； 第三張圖不容易判斷為蘋果鏈格孢葉斑病， |
| ?? Due to the two important issues demonstrated above, we need to use deep learning method to resolve this problem. For obtaining a stable identification result, dataset construction is the most important thing, especially for the training data. For this purpose, we have invited agriculture experts from Chinese Academy of Agricultural Sciences to collect enough effective dataset. In this paper, the agriculture experts have collected apple leaf diseases data from Shaanxi Province, China. In order to well study the identification method, the agriculture experts have mainly collected two kinds of apple leaf diseases: Alternaria leaf blotch and rust. Meanwhile, according to the shape, color of the leaves, the number of diseases in one leaf, etc., we collected 334 images: including 164 Alternaria leaf blotch and 170 rust. These leaves can help us to construct a stable algorithm for apple leaf diseases identification. | ?? 由于上面展示的兩個重要問題，我們需要使用深度學習方法來解決這個問題， 為了獲得穩定的識別結果，資料集構建是最重要的，尤其是對于訓練資料， 為此，我們邀請了中國農業科學院的農業專家來收集足夠有效的資料集， 在本文中，農業專家收集了中國陜西省的蘋果葉病資料， 為深入研究鑒定方法，農業專家主要收集了蘋果葉斑病和銹病兩大類蘋果葉病害， 同時，根據葉片的形狀、顏色、單葉病害數量等，共采集334幅影像：其中鏈格孢葉斑病164幅，銹病170幅， 這些葉子可以幫助我們構建一個穩定的蘋果葉子病害識別演算法， |
4 The MobileNet model for apple leaf diseases identification
| 原文 | 譯文 |
|---|---|
| ?? In this section, the MobileNet model will be introduced for apple leaf diseases identification. | ?? 在本節中，將介紹用于蘋果葉病識別的 MobileNet 模型， |
| ?? hown in Fig. 2, the core architecture is built on depthwise separable convolutions, which happens to be a form of factorized complexities that factorizes a standard complexity into a depthwise complexity and a 1 × 1 complexity is termed as pointwise complexity. Figure 1 shows depthwise separable convolutions with depthwise and pointwise layers followed by batch normal and ReLU. Meanwhile, the model introduces two simple global hyper-parameters to balance the delay and accuracy effectively. The role of the width multiplier α is to thin a network uniformly at each layer. The resolution multiplier ρ is applied to reduce the size of the input image and the internal representation of every layer by the same multiplier. For a feature map of DF × DF in size, the kernel size is Dk × Dk, the input channel is M, the output channel is N. The total computation CM amount for the core layers of our network can be represented as Eq. 1: | ?? 在圖 2 中，核心架構建立在深度可分離卷積上，這恰好是一種分解復雜度的形式，將標準復雜度分解為深度復雜度，1×1 復雜度被稱為逐點復雜度， 圖 1 顯示了具有深度和逐點層的深度可分離卷積，然后是批量法線和 ReLU， 同時，該模型引入了兩個簡單的全域超引數來有效平衡延遲和準確性， 寬度乘數 α 的作用是在每一層均勻地減薄網路， 解析度乘數 ρ 用于通過相同的乘數減小輸入影像的大小和每一層的內部表示， 對于大小為 DF × DF 的特征圖，內核大小為 Dk × Dk，輸入通道為 M，輸出通道為 N，我們網路核心層的總計算量 CM 可以表示為 Eq. 1： |
| 原文 | 譯文 |
|---|---|
| ?? Here, in our experiment for apple leaf diseases identification, we set width multiplier α = 1 and resolution multiplier ρ = 1. In the same situation, the computational cost CS of the standard convolutions can be obtained by Eq. 2: | ?? Here, in our experiment for apple leaf diseases identification, we set width multiplier α = 1 and resolution multiplier ρ = 1. In the same situation, the computational cost CS of the standard convolutions can be obtained by Eq. 2: |
| 原文 | 譯文 |
|---|---|
| ?? Here, in our experiment for apple leaf diseases identification, we set width multiplier α = 1 and resolution multiplier ρ = 1. In the same situation, the computational cost CS of the standard convolutions can be obtained by Eq. 2: | ?? 最后，為了將標準卷積表示為深度卷積和逐點卷積，我們方法中的減少 R 可以通過等式 3 計算： |
| 原文 | 譯文 |
|---|---|
| ?? The most important merit of this model is that it introduces two simple global hyper-parameters to balance the delay and accuracy effectively. These two hyper-parameters, width multiplier and resolution multiplier, allow the model builder to choose the right size model for the application according to the constraints of the problem. | ?? 該模型最重要的優點是它引入了兩個簡單的全域超引數來有效平衡延遲和準確性， 這兩個超引數，寬度乘數和解析度乘數，允許模型構建器根據問題的約束為應用程式選擇合適大小的模型， |
| 原文 | 譯文 |
| :– | :– |
| ?? The most important merit of this model is that it introduces two simple global hyper-parameters to balance the delay and accuracy effectively. These two hyper-parameters, width multiplier and resolution multiplier, allow the model builder to choose the right size model for the application according to the constraints of the problem. | ?? 該模型最重要的優點是它引入了兩個簡單的全域超引數來有效平衡延遲和準確性， 這兩個超引數，寬度乘數和解析度乘數，允許模型構建器根據問題的約束為應用程式選擇合適大小的模型， |
5 The high precision models for apple leaf diseases identification
| 原文 | 譯文 |
|---|---|
| ?? Actually, for obtaining an optimal solution for apple leaf diseases identification, we have tried several deep learning models. The main objectives include high precision and high efficiency. Of course, the balance of these two objectives is necessary. The MobileNet model introduced in previous section can be executed with high efficiency. On the other hand, in order to get a high inspection precision, we have tried ResNet152 [34] and InceptionV3 [35] for apple leaf diseases identification. In this section, we will simply introduce how to inspect apple leaf diseases using ResNet152 and InceptionV3. | ?? 實際上，為了獲得蘋果葉片病害識別的最優解，我們嘗試了幾種深度學習模型， 主要目標包括高精度和高效率， 當然，這兩個目標的平衡是必要的， 上一節介紹的 MobileNet 模型可以高效執行， 另一方面，為了獲得較高的檢測精度，我們嘗試了 ResNet152 [34] 和 InceptionV3 [35] 進行蘋果葉片病害識別， 在本節中，我們將簡單介紹如何使用 ResNet152 和 InceptionV3 檢測蘋果葉片病害， |
5.1 The ResNet152 model for apple leaf diseases identification
| 原文 | 譯文 |
|---|---|
| ?? ResNet was first proposed in 2015, the basic structure of this model consists of two convolutional layers. Meanwhile, this model skips blocks of convolutional layers by using a nonparameterized shortcut connection and adds new inputs into the network and generates new outputs. ResNet were proposed as a family of multiple deep neural networks with similar structures but different depths, i.e. ResNet18, ResNet34, ResNet50, ResNet101 and ResNet152. | ?? ResNet 于 2015 年首次提出，該模型的基本結構由兩個卷積層組成， 同時，該模型通過使用非引數化快捷連接跳過卷積層塊，并將新輸入添加到網路中并生成新輸出， ResNet 被提議為具有相似結構但不同深度的多個深度神經網路家族，即 ResNet18、ResNet34、ResNet50、ResNet101 和 ResNet152， |
| ?? ResNet 于 2015 年首次提出，該模型的基本結構由兩個卷積層組成， 同時，該模型通過使用非引數化快捷連接跳過卷積層塊，并將新輸入添加到網路中并生成新輸出， ResNet 被提議為具有相似結構但不同深度的多個深度神經網路家族，即 ResNet18、ResNet34、ResNet50、ResNet101 和 ResNet152， | ?? 在本文中，我們選擇 ResNet152 作為比較模型，因為它在 ResNet 系列中精度最高 [30]， 如圖 3 所示，一個有病害的蘋果葉子被輸入到 ResNet152 模型中， 四列分別有10層、24層、108層、10層， 總共有152層， 在這里，虛線快捷方式增加了維度，虛線框內的層代表 n 個重復的卷積層， 例如，“虛線框 x 5”表示 5 個重復的卷積層， |
5.2 The InceptionV3 model for apple leaf diseases inspection
| 原文 | 譯文 |
|---|---|
| ?? Inception is a module in GoogleNet. This model can perform multiple convolution operations or pooling operations on the input images in parallel and composing all the results into a very deep feature map. Therefore, its precision should be higher than general models. | ?? Inception 是 GoogleNet 中的一個模塊， 該模型可以對輸入影像并行執行多個卷積操作或池化操作，并將所有結果組合成一個非常深的特征圖， 因此，其精度應高于一般型號， |
| ?? The most effective model of Inception is Inceptionv3 architecture, which is the winner of 2014 ILSVRC, and has 44 layers with 21 million learnable parameters. As shown in Fig. 4, it is the model of InceptionV3 consists of 11 Inception modules. (a) is used in the 1th - 4th and the 10th -11th Inception modules of InceptionV3. (b) is used in the 5th - 9th Inception modules of InceptionV3. However, it is also a time-consuming model. In order to improve this problem, factorizing convolutions were proposed to reduce its parameters. For example, a 5 × 5 filter convolution can be decomposed into two 3 × 3 filter convolutions. Through this step, the parameters in this process reduces from 5 × 5 = 25 to 3 × 3 + 3 × 3 = 18. Thus, it brings 28% reduction in number of parameters. However, the calculation cost is still not able to full satisfy the requirements of efficiency for apple leaf diseases identification. The detail results of efficiency and precision of these models will be shown in Section 6. | ??Inception 最有效的模型是 Inceptionv3 架構，它是 2014 年 ILSVRC 的冠軍，有 44 層，2100 萬個可學習引數，如圖 4 所示，是 InceptionV3 的模型，由 11 個 Inception 模塊組成， (a) 用于 InceptionV3 的第 1 ～ 4 和第 10 ～ 第 11 的 Inception 模塊， (b) 用于 InceptionV3 的第 5 至第 9 個 Inception 模塊，然而，它也是一個耗時的模型，為了改善這個問題，提出了分解卷積來減少其引數，例如，一個 5 × 5 的濾波器卷積可以分解為兩個 3 × 3 的濾波器卷積，通過這一步，這個程序中的引數從5×5=25減少到3×3+3×3=18，從而使引數數量減少了28%，然而，計算成本仍不能完全滿足蘋果葉片病害識別效率的要求，這些模型的效率和精度的詳細結果將在第 6 節中顯示， |
6 Experiments and results
6.1 Experimental setup
| 原文 | 譯文 |
|---|---|
| ?? This experiment was performed on an Ubuntu server (16.04 LTS) with an Intel ? Core? i7-9700KF CPU @ 3.60GHz that was accelerated by an NVIDIA GeForce RTX 2080Ti GPU. NVIDIA GeForce RTX 2080Ti has 4352 CUDA cores and 11 GB memory. The core frequency is up to 1545 MHz. All the deep learning models used in this paper were implemented in the Tensorflow deep learning framework. | ?? 該實驗是在配備 Intel ? Core? i7-9700KF CPU @ 3.60GHz 的 Ubuntu 服務器 (16.04 LTS) 上進行的，該 CPU 由 NVIDIA GeForce RTX 2080Ti GPU 加速， NVIDIA GeForce RTX 2080Ti 擁有 4352 個 CUDA 核心和 11 GB 記憶體， 核心頻率高達 1545 MHz， 本文中使用的所有深度學習模型都是在 Tensorflow 深度學習框架中實作的， |
6.2 Dataset
| 原文 | 譯文 |
|---|---|
?? A dataset containing 334 images were used for training and testing the MobileNet model. All of the images were collected by the agriculture experts who are visiting and surveying kinds of orchards in shaanxi Province. The dataset covers 2 common types of apple leaf diseases: Alternaria leaf blotch (caused by Alternaria alternata f.sp. mali) and rust (caused by Pucciniaceae glue rust). | ?? 包含 334 張影像的資料集用于訓練和測驗 MobileNet 模型， 所有圖片均由正在陜西省各類果園走訪調研的農業專家采集， 該資料集涵蓋了 2 種常見的蘋果葉病害型別：鏈格孢葉斑病（由 Alternaria alternata f.sp. mali 引起）和銹病（由 Pucciniaceae 膠銹病引起）， |
| ?? In order to improve the precision of apple leaf diseases identification, an image generator object was created to perform random rotation, cutting, and grayscale on this dataset. By the above methods, the dataset is expanded. Meanwhile the identification precision will be greatly improved, especially in the case that the photos are taken in different rotation or scaling. Figure 5 shows representative images of the diseased apple leaves in the dataset. In total, the apple leaf dataset contains 2004 images. To perform the experiment, 75% of the dataset is used for training and the other 25% for testing. The ratio of the size of the training dataset to that of the validation dataset is 3:1. Table 1 lists the numbers of training sets and testing sets for apple leaf diseases identification. | ?? 為了提高蘋果葉片病害識別的精度，創建了一個影像生成器物件，對這個資料集進行隨機旋轉、切割和灰度處理， 通過上述方法，擴展了資料集， 同時識別精度將大大提高，尤其是在不同旋轉或縮放比例的情況下， 圖 5 顯示了資料集中患病蘋果葉子的代表性影像， 蘋果葉資料集總共包含 2004 張影像， 為了執行實驗，資料集的 75% 用于訓練，另外 25% 用于測驗， 訓練資料集的大小與驗證資料集的大小之比為 3:1， 表1列出了蘋果葉片病害識別的訓練集和測驗集的數量， |
6.3 Experiment results and analysis
| 原文 | 譯文 |
|---|---|
| ?? In this section, we will give the testing results using three kinds of deep learning models, including MobileNet, InceptionV3, and ResNet152. The results will be demonstrated through the accuracy of convolution neural networks, and average handling time for each image. | ?? 在本節中，我們將使用 MobileNet、InceptionV3 和 ResNet152 三種深度學習模型給出測驗結果， 結果將通過卷積神經網路的準確性和每張影像的平均處理時間來證明， |
| ?? The average handing times are used to evaluate efficiency for apple leaf diseases identification. As shown in Table 2, MobileNet based method is the most efficient one, only 0.22 s is taken for each image. However, the apple leaf diseases identification time for the InceptionV3 model is more than 2 times than that of MobileNet model. Furthermore, the cost of ResNet152 is nearly 4 times than that of MobileNet model. Therefore, the efficiency of MobileNet model is the best one. | ??平均處理時間用于評估蘋果葉片病害識別的效率， 如表 2 所示，基于 MobileNet 的方法是最有效的方法，每張影像只需要 0.22 秒， 然而，InceptionV3 模型的蘋果葉病害識別時間是 MobileNet 模型的 2 倍以上， 此外，ResNet152 的成本是 MobileNet 模型的近 4 倍， 因此，MobileNet 模型的效率是最好的， |
| 原文 | 譯文 |
|---|---|
| ?? Finally, the accuracy of these 3 models are nearly the same. Supposing the testing accuracy is represented by Atesting. It can be obtained by calculating the ratio between the number of corrected identified (Ncorrect) and the number of the total images (Ntotal), as Eq. 4. | ?? 最后，這 3 個模型的準確率幾乎相同， 假設測驗精度用Atesting表示， 可以通過計算校正識別數（Ncorrect）與總影像數（Ntotal）的比值得到，如式4所示， |
7 Conclusions and future works
| 原文 | 譯文 |
|---|---|
| ?? In this paper, a MobileNet based apple leaf diseases identification method is proposed. This method can greatly reduce the burden of experts for apple leaf diseases identification. It can offer a stable identification result. Meanwhile, it is a low-cost method, because it can be easily deployed in a mobile device. Furthermore, we also provide a good balance between the efficiency and precision, this is achieved through comparing several deep learning models. | ?? 本文提出了一種基于MobileNet的蘋果葉病害識別方法， 這種方法可以大大減輕專家對蘋果葉片病害鑒定的負擔， 它可以提供穩定的識別結果， 同時，它是一種低成本的方法，因為它可以很容易地部署在移動設備中， 此外，我們還在效率和精度之間提供了良好的平衡，這是通過比較幾種深度學習模型來實作的， |
| ?? One of possible extension is to collect more datasets for further improving the identification precision. In our future plan, we are going to collect 2,000,000 images for all apple leaf diseases as a training dataset. This will greatly help us to develop a much more effective deep learning models for apple leaf diseases identification. | ?? 一種可能的擴展是收集更多資料集以進一步提高識別精度， 在我們未來的計劃中，我們將收集 2,000,000 張所有蘋果葉病害的影像作為訓練資料集， 這將極大地幫助我們開發更有效的蘋果葉片病害識別深度學習模型， |
| ?? Another future work is to design a deep learning model to quality the apple leaf diseases. For example, we can give the different levels of two rust in Fig. 1. This is a challenge task, but it is an urgent requirement from agricultural experts. | ?? 未來的另一項作業是設計一個深度學習模型來質量蘋果葉病， 例如，我們可以在圖 1 中給出兩種銹病的不同程度，這是一項具有挑戰性的任務，但卻是農業專家的迫切要求， |
轉載請註明出處，本文鏈接：https://www.uj5u.com/qita/293445.html
標籤：其他
上一篇：2021-08-11
下一篇：ubuntu20.04系統+cuda10.1+opencv4.1.0+dense_flow的編譯與安裝，吐血總結！！！