Hiperspektral Görüntülerde Derin Öğrenme ile Hedef Tespiti
Severoğlu, Batuhan Mert
xmlui.mirage2.itemSummaryView.MetaDataShow full item record
In recent years, hazardous chemicals such as explosives and narcotics have caused many people and particularly security forces to experience major problems in various areas. This has made the detection of hazardous chemicals as early as possible a critical task. Several systems and algorithms using different types of data have been developed to fulfill this task. However, in the last period, the increase in the diversity of hazardous chemicals, transportation methods and the use of trace amounts together with the developing technology caused the existing algorithms to fall short. Therefore, studies have been initiated on systems that offer more comprehensive solutions for the detection of chemical substances. Since the detection of chemicals is a target detection task, the Hyperspectral Imaging (HSI) technique, which is frequently used in many areas of target detection, has become an important alternative for the detection of chemicals. Furthermore, deep learning models such as Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM), which have proven their success in many fields in recent years, are among the most used methods of target detection. In the scope of this thesis, hyperspectral images were collected by placing four different chemical substances on four different background materials with a mechanism using Hyperspectral Imaging. A separate dataset for each chemical was formed with pixels selected by certain methods in these images. These datasets have been used in several CNN and LSTM models for chemical substance detection. The data were first passed through the Differential Reflectometry (DR) process during the pre-processing stage. DR is the process of normalizing the difference between the reflection values of two sequential scans. The attributes in the spectral dimension of the hyperspectral data collected by this differential-like process are made distinctive. In the second part of the pre-processing stage, Savitzky-Golay filter was applied to the spectral dimension of the hyperspectral data with the intent of feeding smooth and noise-free spectral data to the deep learning models. Thus, spectral data prepared in this way are fed into deep learning models. In the scope of this thesis, three types of deep learning model have been discussed. Firstly, one-dimensional CNN model was discussed. The inputs of this model are one-dimensional spectral vectors and all processes in the model are one-dimensional. In the two-dimensional CNN model developed later, the input data and all the processes in the model are two-dimensional operations. The matrix-shaped inputs of the two-dimensional CNN model were obtained by reshaping one-dimensional spectral data. In the lastly developed LSTM model, one-dimensional spectral information was used directly. Target detection performances of four different chemical substances of the developed models on four different background materials were investigated. Situations that critically affect the detection of chemical targets are considered. These cases are the type of chemical substance, the type of background material, the number of availability of data in the dataset, the amount of the chemical on the surface, the shape and edges of objects in the image. In addition, the advantages and disadvantages of the methods tried to increase the performance of the model in the post-processing, training and final processing stages are exhibited. As a result, it was seen that the CNN models had better performance than the LSTM models for the used dataset. The improved CNN models demonstrate high performance in the detection of most chemicals. It is also stated that the transition from the one-dimensional CNN model to the two-dimensional CNN model has significantly improved the performance by making the data two-dimensional. With the models developed within the scope of this thesis, it has been understood that hazardous chemical substances can be detected successfully with CNN models, and it has been seen that the LSTM models are also promising for this task.