Homojen olmayan bir yüzey altında gömülü nesneler için evrişimsel sinir ağı tabanlı hedef tespiti
Convolutional neural network based target detection for objects buried under a non-homogeneous surface
- Tez No: 954353
- Danışmanlar: PROF. DR. ALİ YAPAR
- Tez Türü: Yüksek Lisans
- Konular: Elektrik ve Elektronik Mühendisliği, Savunma ve Savunma Teknolojileri, Electrical and Electronics Engineering, Defense and Defense Technologies
- Anahtar Kelimeler: Belirtilmemiş.
- Yıl: 2025
- Dil: Türkçe
- Üniversite: İstanbul Teknik Üniversitesi
- Enstitü: Lisansüstü Eğitim Enstitüsü
- Ana Bilim Dalı: Savunma Teknolojileri Ana Bilim Dalı
- Bilim Dalı: Savunma Teknolojileri Bilim Dalı
- Sayfa Sayısı: 148
Özet
Yeraltı yapılarının ve gömülü nesnelerin tespiti, mühendislik, arkeoloji ve savunma sanayii gibi birçok alanda kritik öneme sahiptir. Bu amaçla yaygın olarak kullanılan Yer Radarı (GPR), yeraltına gönderilen elektromanyetik dalgaların yansımalarını analiz ederek hedeflerin tespiti için tahribatsız bir yöntem sunar. Ancak, ham GPR verileri genellikle karmaşık yansımalar, yüzey dağınıklığı, gürültü ve zayıf hedef sinyalleri içerdiğinden doğrudan analize uygun değildir. Bu tez çalışmasında, sentetik GPR B-tarama verilerinden yeraltında gömülü nesnelerin (özellikle mayın benzeri hedeflerin) otomatik tespiti için, optimize edilmiş hibrit bir ön işleme yaklaşımı ve Evrişimsel Sinir Ağı (CNN) tabanlı bir sınıflandırma modeli geliştirilmiştir. Çalışmada, gprMax yazılımı kullanılarak tek katmanlıdan dört katmanlıya kadar farklı zemin ortamlarında, çeşitli hedef malzemeleri (PEC, plastik, plastik-PEC karışımı, ahşap), farklı derinlik, boyut ve anten konumlarını içeren 2880 adet sentetik B-tarama verisi oluşturulmuştur. Bu ham veriler, hedef sinyallerini belirginleştirmek ve CNN modelinin öğrenme performansını artırmak amacıyla kapsamlı bir ön işleme sürecinden geçirilmiştir. Uygulanan hibrit ön işleme zinciri; Gaussian filtre tabanlı arka plan çıkarma, yüksek geçirgen filtreleme, eğrilik artırıcı filtre (CEF), beyazlatma, zaman kazancı ve Non-Local Means (NLM) gürültü azaltma adımlarını içermektedir. Bu adımların parametreleri ve uygulanma sırası, sinyal kalitesini ve hedef belirginliğini maksimize etmek üzere kapsamlı deneysel analizlerle optimize edilmiştir. Ayrıca, modelin genelleme yeteneğini artırmak ve küçük veri setinin neden olduğu aşırı öğrenmeyi azaltmak için eğitim sırasında Gauss gürültüsü ekleme, rastgele dikey kaydırma ve rastgele parlaklık ayarı gibi veri artırma teknikleri kullanılmıştır. Ön işlenmiş ve standardize edilmiş verilerle, kademeli filtre artışı, Batch Normalization, Global Average Pooling ve Dropout katmanları içeren optimize edilmiş bir CNN mimarisi tasarlanmış ve eğitilmiştir. Modelin performansı ve genelleme yeteneği, farklı hiperparametre kombinasyonları ve veri seti büyüklükleri için test edilerek, nihai başarım 5-Katmanlı Çapraz Doğrulama (K-Fold Cross-Validation) yöntemiyle değerlendirilmiştir. Elde edilen sonuçlar, geliştirilen sistemin, artan ortam karmaşıklığına ve veri çeşitliliğine adapte olarak yeraltı hedeflerini %98'in üzerinde bir ortalama doğrulukla başarılı bir şekilde tespit edebildiğini göstermiştir. Özellikle, modelin tüm gerçek hedefleri tespit etme başarısı (Recall) ve pozitif tahminlerindeki kesinliği (Precision) dikkate değer bulunmuştur. Bu çalışma, karmaşık sentetik GPR verilerinde etkili bir ön işleme ve derin öğrenme yaklaşımının, yeraltı hedef tespiti problemlerine umut verici ve yüksek başarımlı çözümler sunabileceğini ortaya koymaktadır.
Özet (Çeviri)
The detection of subsurface structures and buried objects is of critical importance in numerous fields, including engineering, archaeology, and the defense industry. Ground Penetrating Radar (GPR), a widely utilized method for this purpose, offers a non-destructive technique for the detection and characterization of targets by analyzing the reflections of electromagnetic waves transmitted into the subsurface. However, raw GPR data are often unsuitable for direct analysis due to complex reflections, surface clutter, noise, and weak target signals. These challenges make the interpretation of B-scans a task that heavily relies on expert operators, which is often slow, subjective, and prone to error. This thesis addresses these limitations by developing a robust, automated framework for target detection. The primary objective is to create a hybrid preprocessing approach and a Convolutional Neural Network (CNN) based classification model for the automatic detection of buried objects, particularly mine-like targets, from synthetic GPR B-scan data. To build a comprehensive and diverse dataset for training and evaluating the deep learning model, this study utilized gprMax, an open-source FDTD (Finite-Difference Time-Domain) based electromagnetic simulation software. A total of 2880 unique synthetic B-scan datasets were generated, systematically varying multiple key parameters to cover a wide range of realistic scenarios. The variations included: Subsurface Environment Complexity, with models ranging from single homogeneous layers (dry sand, asphalt, stabilized material) to complex multi-layered environments of up to four distinct layers; Target Material Composition, encompassing four types of targets: Perfect Electric Conductors (PEC) to simulate metallic mines, plastic, wood, and a composite plastic-PEC mix to represent complex targets; Target and Antenna Geometry, with variations in target depth, size, horizontal position, and antenna height relative to the surface. This systematic generation of a large-scale, richly annotated dataset is a crucial contribution, enabling a robust evaluation of the proposed methods under controlled yet complex conditions. Raw gprMax B-scans underwent a comprehensive, multi-stage hybrid preprocessing procedure to enhance target signals and improve the learning performance of the CNN model. The development of this chain involved extensive experimental analysis to determine the optimal sequence and parameters for each step. The final implemented hybrid preprocessing chain included: Gaussian Filter-based Background Removal: To suppress low-frequency vertical trends and noise within each A-scan independently. High-Pass Filtering (HPF): Applied to mitigate low-frequency noise and artifacts, such as those arising from the air-ground interface. Curvature Enhanced Filter (CEF): A critical step to specifically amplify the hyperbolic signatures of buried targets. The sigma parameter of the CEF was meticulously optimized, as it was found to be highly sensitive to the target's geometry and burial environment. Whitening: To normalize the signal's amplitude and equalize its frequency spectrum, further enhancing the saliency of target structures. Time Gain: Applied to compensate for the exponential decay of the GPR signal with depth, ensuring that deeper targets are as visible as shallower ones. Non-Local Means (NLM) Noise Reduction: Employed as a final step to remove random noise while preserving the structural integrity of the hyperbolic reflections. The optimization of this chain, particularly the order of operations and the tuning of parameters like the CEF's sigma, was found to be a decisive factor in the model's final performance, highlighting the critical role of tailored preprocessing in GPR data analysis. An optimized CNN architecture was designed to effectively learn the hierarchical features from the preprocessed B-scans. The architecture was built upon modern deep learning principles to ensure high performance and robust generalization, even with a limited dataset. The proposed model features: A sequence of three convolutional blocks with a progressive increase in filter count (16 → 32 → 64). This hierarchical structure allows the model to learn simple features like edges and curves in the initial layers, and then combine them to recognize more abstract and complex patterns, such as the full hyperbolic shape, in deeper layers. Batch Normalization applied after each convolutional layer to stabilize the training process, accelerate convergence, and act as a form of regularization. Global Average Pooling (GAP) instead of a traditional Flatten layer. This step dramatically reduces the number of model parameters, significantly mitigating the risk of overfitting and making the model more robust to spatial translations of the target. A final classification head consisting of a Dense layer, followed by a Dropout layer (with a rate of 0.3) for strong regularization, and a single-neuron output layer with a sigmoid activation function to produce the final probability score for binary (target/non-target) classification. L2 regularization was also applied to convolutional and dense layers to further prevent overfitting. This carefully designed architecture proved to be highly effective for the GPR target detection task. The model's training process was managed by a comprehensive set of strategies to maximize performance and ensure reliable evaluation. Crucially, all data was standardized based only on the statistics (mean and standard deviation) of the training set to prevent data leakage. To enhance the model's generalization capability and mitigate overfitting, several data augmentation techniques were employed during training. These on-the-fly transformations included: Gaussian noise addition to simulate sensor noise, random vertical shifting (RandomTranslation) to mimic variations in the time-zero or surface level, and random brightness adjustment to simulate varying signal strengths. The model was compiled using the Adam optimizer with an initial learning rate of 0.001 and binary_crossentropy as the loss function. The training process was governed by two key callbacks: EarlyStopping, to halt training when the validation loss ceased to improve, and ReduceLROnPlateau, to adaptively decrease the learning rate. The model's final performance and robustness were not evaluated on a single train-test split but through a rigorous 5-Fold Cross-Validation method. This approach provides a more reliable estimate of the model's generalization ability by training and validating it on different subsets of the data. The obtained results demonstrate that the developed system can successfully detect subsurface targets with exceptional performance. The 5-Fold Cross-Validation on a dataset of 864 single-, dual-, and triple-layer environments yielded an average test accuracy of 98.03% and an average F1-Score of 98.11%. Notably, the model achieved an average Precision of 98.43% and an average Recall of 97.79%, indicating that it is highly effective at both avoiding false alarms and detecting nearly all true targets. This study reveals that an effective preprocessing chain, combined with a well-regularized deep learning model and appropriate data augmentation, can offer highly promising and reliable solutions to subsurface target detection problems.
Benzer Tezler
- Eşzamanlı karbondioksit kullanımı ve hidrojen dönüşümü için nikel-demir bazlı katalizör üzerinde karbondioksitten metan eldesi
CO2 to methane over nickel-iron based catalyst for simultaneous CO2 utilization and hydrogen conversion
BUKET KAPLAN
Yüksek Lisans
Türkçe
2023
Kimya Mühendisliğiİstanbul Teknik ÜniversitesiKimya Mühendisliği Ana Bilim Dalı
DOÇ. DR. ALPER SARIOĞLAN
PROF. DR. ŞAHİKA SENA BAYAZİT
- Structural health monitoring of frp composites using fbg sensors
Frp kompozitlerin yapısal sağlık takibinin fbg sensörler kullanılarak gerçekleştirilmesi
ERDEM AKAY
Doktora
İngilizce
2018
Havacılık Mühendisliğiİstanbul Teknik ÜniversitesiUçak ve Uzay Mühendisliği Ana Bilim Dalı
PROF. DR. HALİT SÜLEYMAN TÜRKMEN
PROF. DR. MEHMET YILDIZ
- Size effect on anchorage length of fiber-reinforced concrete structures
Lif katkılı betonların kenetlenme boyu ve boyut etkisi
AHMED ARSLAN
Yüksek Lisans
İngilizce
2024
İnşaat Mühendisliğiİstanbul Teknik Üniversitesiİnşaat Mühendisliği Ana Bilim Dalı
DOÇ. DR. AHMET ABDULLAH DÖNMEZ
- Evaluating the LCA of two buildings with close embodied energy which have different functions
Farklı işlevlere sahip olan iki binanın üç tür duvar kullanarak yaşam döngüsünün değerlendirilmesi
POOYA PAKMEHR
Yüksek Lisans
İngilizce
2014
Mimarlıkİstanbul Teknik ÜniversitesiMimarlık Ana Bilim Dalı
DOÇ. DR. MUSTAFA ERKAN KARAGÜLER
- Microwave imaging of breast cancer with contrast agents
Meme kanserinin kontrast ajanlarla mikrodalga görüntülemesi
SEMA YILDIRIM
Doktora
İngilizce
2020
Elektrik ve Elektronik Mühendisliğiİstanbul Teknik ÜniversitesiElektronik ve Haberleşme Mühendisliği Ana Bilim Dalı
DOÇ. DR. MEHMET ÇAYÖREN