Geri Dön

Makine öğrenmesi ve görüntü işleme tekniklerini kullanarak drone ile yaprak sınıflandırma

Leaf classification with drone by using machine learning and image processing techniques

PDF İndir

  1. Tez No: 512704
  2. Yazar: MEHMET ÖZTÜRK
  3. Danışmanlar: DR. ÖĞR. ÜYESİ NAZIM KEMAL ÜRE
  4. Tez Türü: Yüksek Lisans
  5. Konular: Uçak Mühendisliği, Aircraft Engineering
  6. Anahtar Kelimeler: Belirtilmemiş.
  7. Yıl: 2018
  8. Dil: Türkçe
  9. Üniversite: İstanbul Teknik Üniversitesi
  10. Enstitü: Fen Bilimleri Enstitüsü
  11. Ana Bilim Dalı: Uçak ve Uzay Mühendisliği Ana Bilim Dalı
  12. Bilim Dalı: Uçak ve Uzay Mühendisliği Bilim Dalı
  13. Sayfa Sayısı: 97

Özet

Bu tez çalışmasında makina öğrenmesi ve görüntü işleme tekniklerini kullanarak, drone ile farklı bitki çeşitlerinden ortaya çıkan yaprakların ¸sekilsel olarak sınıflandırılması yapılmıştır. Dünya üzerindeki insan nüfusunun artması ile beraber ortaya çıkan“Hassas Tarım”ile birlikte teknoloji ile tarım daha içe içe olmuştur. Bu tez çalışmasındaki temel motivasyon, insansız hava araçları vasıtasıyla otonom bir şekilde yaprak çeşitlerini tanıyarak tarımda kimyasal ve biyolojik iyileşmenin önünü açmaktır. İlaveten, drone görüntülemesi sayesinde bitkisel hastalıklar ve böceklenme gibi ortaya çıkabiliecek sorunlar daha önceden fark edilerek gerekli tedbirlerin alınmasını kolaylaştırmaktadır. Bu çalışma içerisinde yer alan metotlar kullanılarak insansız hava araçları sayesinde taranan bölgelerin bitki veri tabanı oluşturulabilir, bu sayede daha sürdürebilir bir bitki ekosistemi kurulabilir. Temel olarak çalışma 4 bölüme ayrılmıştır. 1) Literatür taraması yapılarak bu alanda yapılmış çalışmalar incelenmiştir. Böylece, bu tezde yer alması hedeflenen çalışmalara sıfırdan başlanıp mevcut literatür ile tekrar düşmektense, var olan çalışmaların eksik/yetersiz olduğu yönler tespit edilerek bu konuların üzerine gidilmiştir. İlaveten, bu çalışmalar sırasında Flavia, imageclef datasetlerinin yanında, çeşitli mesafelerden kendi drone ile çektiğimiz görüntü dataları da kullanılmıştır. 2) Elde bulunan veritabanındaki görsellerde yaprakların resimleri arkaplanda bulunan ambiyans ile bir bütündü çalışmanın başında. Lakin, sınıflandırmanın doğru yapılması için, yaprak görsellerinin arkaplanda gürültü unsuru olabilecek görselliklerden ayrılması gerekmektedir. Bu nedenle, tüm görsellerde yeşil olmayan arkaplan çıkartılmıştır. 3) Drone ile daha uzak mesafeden görüntü alınarak uzaktan yapılan çekimlerin sonucudaki başarı oranı sınanmıştır. Bu bölümdeki başarı oranını arttırmak amacıyla RDP doygunluk (saturation) ve Parlaklık (value) eşitlemesi gibi teknikler uygulanmıştır. 4) Üst üste binmiş yapraklar üzerinden elde edilen yaprak görünütülerinin sınıflandırılması yapılmıştır. Daha önceki bölümlerde arka planda yaprak harici nesnelerin bulunduğu durumlar incelenmiştir. Bu bölümde ise, arka plandaki nesneler içinde yaprakların da olabileceği görseller incelenmiştir. Bu tarz durumlardaki sınıflandırmaların başarıya ulaşması için Watershed algoritması uygulanmıştır. Her bölümde de, temel olarak görüntünün işlenmesi, özelliklerin cıkartılması, özelliklerin öğretilerek SVM ile sınıflandırma yapılmıştır. Çalışmanın sonunda, insansız hava aracı ile belirli bir bölgede görüntü işleme ve makine öğrenmesi ile yapılacak olan yaprak sınıflandırmasının %91,3 oranı ile başarıya ulaşacağı bir sistem tasarlanmıştır.

Özet (Çeviri)

In this thesis study, classification of different tree leaves are achieved through machine learning and image processing techniques. The study's experimental aspect is conducted using drones to capture images from multiple aerial points of view. The primary motivation of this thesis is the stabilization of ecological floral balance and the progression of precision agriculture. Plants have a major role on ecology on Earth and are the fundamental basis of the energy pyramid. Henceforth, the technological involvement on the sustainability of plants is vital as the human population increases exponentially. A central methodology of achieving such goal relies on autonomy, and a widespread choice of autonomous vehicles are drones due to their cheap cost and easy maintenance. This study, thus, focuses on the application of drones on such appliances. This thesis targets the recognition and classification of various leaf shapes for the purpose of paving the way for chemical and biological developments in agriculture, as well as preventing detrimental effects on plants due to illnesses and bug/pest contamination. In addition, with the help of drone footage captures, a floral database of plants may be constructed, which in return, supports a progression towards a more sustainable ecological balance. In existing literature, there does not exist any direct work on the classification of tree plants using machine learning and image processing. However, there do exist studies on photographs taken from cameras. Some of these studies include: Flavia leaf database, neural network leaf classification from leaf vessels, multi-path curvelet transformation, et cetera. Since the application of drones is similar to static camera approaches, previous studies can be used as to find a starting basis for the work comprised in this thesis. However, the application of drones and autonomous categorization is not found in existing literature. This study contributes to both the spread of drone use and the support to agriculture and farmers. On the other hand, there exists multiple challenges, some being: the angle and altitude of the quad-copter, distance to leaf, vibrations inflicted on of footage capture device, luminous intensity. The studies in this thesis can be divided into 4 parts: 1) Literature survey is conducted to gain an understanding of previous literature. By doing so, the techniques aimed to take place in this thesis can focus on the drawbacks of previously existing literature by benefiting from the methodologies presented in them and consequently building the work of this thesis on them. This approach both prevents the repetition of existing literature and enables the improvement over previous work. In addition, this study uses drone captures along with imageclef and Flavia datasets, which permit a much more wider range of leaf types and photo angles to study on. Though, before the use of Flavia data, preprocessing to utilize the most important parameters of the images have been conducted. That being said, image processing has its own difficulties, mainly due to: i) data loss from the transition of 2D from 3D, ii) the necessity of producing a different mathematical model for each and every object, iii) excessive amount of noise in the real world, iv) storage problems of multimedia data, v) number of pixel and luminous effects captured. To counter such problems, multiple properties of photographs are to be modified before classification. The RGB is the first of these, and signifies the combination of red, green and blue pixels on the photograph in a numerical way. The second is the intensity, which is a single matrix signifying the closeness of a pixel to absolute white of absolute black. The properties identified for the machine learning algorithm to use are defines as follows: I) The diameter of the leaf, which can be measures as the distance between the farthest two points on the leaf. This distance is found after the Laplacian and average filtering, where the distance is in terms of Euclidean norm. A Matlab algorithm using regionprobs is also used, which returns measurements for the set of properties specified by properties for each 8-connected component (object) in the binary image. II) Physiologic distance, which coincides to the distance between two points on the leaf main vessel. This distance has to be input manually under normal circumstances. However, the Matlab function of regionprobs was used to automatize to process. III) Physiologic width, which are orthogonal lengths to the physiologic distance. To automatize and simplify the interpretation, the short diameter of the leaf is taken into account. IV) Leaf area, which is the amount of white pixels. V) Perimeter, which is the amount of pixels covering the leaf margin. 2) The database from Flavia consists of photographs which have absolute white backgrounds behind the leaves. However, in actual footage obtained from drones, it is practically impossible to obtain photographs with clear backgrounds. And so, it is vital to first clear any irrelevant objects in the background before attempting to classify the leaves. This section covers the cases where the background only consists of non-leaf material, where in short, the approach removes all of the non-green toned objects in the image. By doing so, the leaf object is extracted ftom the image's irrelevant background objects. The methodology for solving such problem was achieved via a technique called Thresholding, which is the simplest form of segmentation of an image. If the red and green values are increased in the RGB channel, then the extraction of the lead can be achieved. Experimental studies conducted to verify this approach have shown that the leaf object is easily extracted from the background when red and green channels are intensified. 3) This part of thesis examines the success rate of long-distance drone footage captures. Shots from 2, 4, 7 and 11 meters were taken via a drone camera. Algorithms developed in the previous sections were put into trial. However, major reasons are due to the distance of leaves from the camera, the leaves were mostly not recognized and an intolerable failure surfaced. It was also seen that besides the factor of distance between leaves and the drone, luminous intensity, color and contrast also has a major effect. The HSV (hue-saturation-value) has been modified in the purpose to increase the success rate of leaf recognition from photographs. When the saturation and the value histogram equalizations are performed, the leaf was successfully extracted from the image. However, this caused other irrelevant objects in the background with green color to be extracted as well. The value of the HSV was transformed to histogram equalizations for Matlab to correctly interpret, but an even better solution turned out to be the imadjust method where differecent contrast values at tried. The contrast reduction technique is similar to histogram equalization, however, as opposition to histogram equalization, contrast reduction equals values beneath average to minimum, and above average to maximum. This causes a binary outcome, which is not handy in application. This surfaces the necessity to manually remove all other irrelevant objects in the background. 4) This section of the thesis covers the cases where different types of leaves are present in the background of the primary leaf. For images shot from a single angle, difficulties arise for determining the depth of the leaves. To counter this difficulty, a method called Watershed has been implemented. The concept of Watershed relies on projecting a 3D image on a 2D space with intensity image. This is achieved by finding three points: regional minimum, water-drop points (where water will definitely flow), and equal water level points (flow to multiple minimums). The second of these conditions are catchment basins and the third of these conditions are divide lines. The main purpose of this concept is to find the watershed lines. The intensity component becomes the main property in distinguishing the primary leaf from other leaves. For each of the four parts, the main application is the processing of the images, extraction of dominant properties, learning the properties and categorization using SVM. For the first part, the results show that the success rate is 91.3% compared to previous studies. The necessity for manual interference has been overcome and the process has been automatized. The the second part where the non-green background is removed, the study does not show any significant difference with respect to distance from leaves. For the third part where long-distant captures are classified, the error for such cases are less than 1%. For the last part where leaves on top of each other are examined, the segmentation is divided into three titles: good, bad, failed. A good segmentation is where the extracted leaf is very close to the primary leaf shape in the actual photo. A bad segmentation is where incomplete extraction is present from the photo. A failed segmentation is is irrelevant background images are present in the extraction and the extracted leaf image is unusable. The main purpose of this thesis was the classification of leaf images obtained from drones using machine learning, and the automation of this process. This, in return, will elevate the use of real-time applications in agriculture while decreasing human intervention to the process of plant classifications. The results of such developments are smart environmental surveillance of flora, precision agriculture, and furthermore, the proper determination of plant ecology on Earth. Another benefit is the use of such system where humans are incapable of entering, such as forests and mountains where accessibility to such area would be possible for a real-time autonomous scan of the terrestrial texture of such inhabitable locations. The experimental studies taken place in this thesis has been conducted in Control and Avionics Laboratory and Aerospace Research Center in Istanbul Technical University, where the flights of the drone, of which captured photos have been used as a dataset, have been performed by a trained pilot.

Benzer Tezler

  1. Tez No

    688973

    Deep feature transfer from deep learning models into machine learning algorithms to classify COVID-19 from chest X-ray images

    Göğüs röntgeni görüntülerinden COVID-19 sınıflandırması yapmak amacıyla derin öğrenme modellerinden makine öğrenmesi algoritmalarına derin öznitelik aktarımı

    OZAN GÜLDALİ

    Yüksek Lisans

    İngilizce

    İngilizce

    2021

    Matematikİstanbul Teknik Üniversitesi

    Matematik Mühendisliği Ana Bilim Dalı

    DR. ÖĞR. ÜYESİ GÜL İNAN

  2. Tez No

    849108

    DA4HI: A deep learning framework for facial emotion recognition in affective systems for children with hearing impairments.

    DA4HI: İşitme engelli çocuklar için duyuşsal sistemlerde yüzdeki duyguların tanınması maksadıyla geliştirilen derin öğrenme modeli.

    CEMAL GÜRPINAR

    Doktora

    İngilizce

    İngilizce

    2023

    Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrolİstanbul Teknik Üniversitesi

    Bilgisayar Mühendisliği Ana Bilim Dalı

    PROF. DR. HATİCE KÖSE

    PROF. DR. NAFİZ ARICA

  3. Tez No

    717024

    Satellite images super resolution using generative adversarial networks

    Uydu görüntülerinde çekişmeli üretici ağ kullanarak süper çözünürlük

    MARYAM SERDAR

    Yüksek Lisans

    İngilizce

    İngilizce

    2022

    Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrolİstanbul Teknik Üniversitesi

    İletişim Sistemleri Ana Bilim Dalı

    PROF. DR. AHMET HAMDİ KAYRAN

  4. Tez No

    883275

    Investigation of microstructure movement under flow by using image processing and deep learning

    Akış altındaki mikroyapı deformasyonunun görüntü işleme ve derin öğrenme kullanılarak incelenmesi

    SAEED SARBAZZADEH KHOSROSHAHI

    Yüksek Lisans

    İngilizce

    İngilizce

    2023

    Elektrik ve Elektronik Mühendisliğiİstanbul Teknik Üniversitesi

    Elektronik ve Haberleşme Mühendisliği Ana Bilim Dalı

    DR. ÖĞR. ÜYESİ AHMET CAN ERTEN

  5. Tez No

    847190

    Prediction of COVID 19 disease using chest X-ray images based on deep learning

    Derin öğrenmeye dayalı göğüs röntgen görüntüleri kullanarak COVID 19 hastalığının tahmini

    ISMAEL ABDULLAH MOHAMMED AL-RAWE

    Yüksek Lisans

    İngilizce

    İngilizce

    2024

    Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve KontrolGazi Üniversitesi

    Bilgisayar Mühendisliği Ana Bilim Dalı

    PROF. DR. ADEM TEKEREK

Geri Dön