Browsing by Subject "Neural network"

Filter results by typing the first few letters
Now showing 1 - 5 of 5
  • Thumbnail Image
    ItemOpen Access
    Additive neural network for forest fire detection
    (Springer, 2020) Pan, H.; Badawi, D.; Zhang, X.; Çetin, Ahmet Enis
    In this paper, we introduce a video-based wildfire detection scheme based on a computationally efficient additive deep neural network, which we call AddNet. This AddNet is based on a multiplication-free vector operator, which performs only addition and sign manipulation operations. In this regard, we construct a dot product-like operation from the mf-operator and use it to define dense and convolutional feed-forwarding passes in AddNet. We train AddNet on images taken from forestry surveillance cameras. Our experiments show that AddNet can achieve a time-saving by 12.4% when compared to an equivalent regular convolutional neural network (CNN). Furthermore, the smoke recognition performance of AddNet is as good as regular CNNs and substantially better than binary-weight neural networks.
  • Thumbnail Image
    ItemOpen Access
    Feasibility of impact-acoustic emissions for detection of damaged wheat kernels
    (Elsevier BV, 2007-05) Pearson, T.; Çetin, A. Enis; Tewfik, A. H.; Haff, R. P.
    A non-destructive, real time device was developed to detect insect damage, sprout damage, and scab damage in kernels of wheat. Kernels are impacted onto a steel plate and the resulting acoustic signal analyzed to detect damage. The acoustic signal was processed using four different methods: modeling of the signal in the time-domain, computing time-domain signal variances and maximums in short-time windows, analysis of the frequency spectrum magnitudes, and analysis of a derivative spectrum. Features were used as inputs to a stepwise discriminant analysis routine, which selected a small subset of features for accurate classification using a neural network. For a network presented with only insect damaged kernels (IDK) with exit holes and undamaged kernels, 87% of the former and 98% of the latter were correctly classified. It was also possible to distinguish undamaged, IDK, sprout-damaged, and scab-damaged kernels.
  • Thumbnail Image
    ItemOpen Access
    Learning based control compensation for multi-axis gimbal systems using inverse and forward dynamics
    (2021-09) Leblebicioğlu, Damla
    Unmanned aerospace vehicles (such as rockets, drones, and satellites) usually carry sensors as their primary payload. These sensors (i.e., electro-optical and/or infrared imaging cameras) are used for image processing, target tracking, surveillance, mapping, and providing high-resolution imagery for environmental surveys. It is crucial to obtain a steady image in all of those applications. This is typically accomplished by using multi-axis gimbal systems. This study concentrates on the modeling and control of a multi-axis gimbal system that will be mounted on a surface-to-surface tactical missile. A novel and fully detailed procedure is proposed to derive the nonlinear and highly coupled EOMs (Equations of Motion) of the two-axis gimbal system. Different from the existing works, Forward Dynamics of the two-axis gimbal system is modeled using multi-body dynamics modeling techniques. In addition to Forward Dynamics model, Inverse Dynamics model is generated to estimate the complementary torques associated with the state and mechanism-dependent, complex disturbances acting on the system. Forward and Inverse Dynamics models are used in Monte Carlo Simulations (MCSs) for Sensitivity Analysis. A multilayer perceptron (MLP) structure based disturbance compensator is implemented to cope with external and internal disturbances and parameter uncertainities through torque input channel. Comparisons with well known controllers such as cascaded PID, ADRC (Active Disturbance Rejection Control), Inverse Dynamics based controllers show that the NN (neural network)-based controller is more succesful in the full operational range without requiring any tuning or adjustment. Implementation of MLP assisted closed-loop control with simulations using Simulink® are performed. Finally, proposed control algorithms are tested on the physical system by using Simulink® Real-Time (xPC Target). Comparative results are presented in figures and tables in the thesis.
  • Thumbnail Image
    ItemOpen Access
    Malignant-lesion segmentation using 4D co-occurrence texture analysis applied to dynamic contrast-enhanced magnetic resonance breast image data
    (2007) Woods, B.J.; Clymer, B.D.; Kurc, T.; Heverhagen J.T.; Stevens, R.; Orsdemir, A.; Bulan O.; Knopp, M.V.
    Purpose: To investigate the use of four-dimensional (4D) co-occurrence-based texture analysis to distinguish between nonmalignant and malignant tissues in dynamic contrast-enhanced (DCE) MR images. Materials and Methods: 4D texture analysis was performedon DCE-MRI data sets of breast lesions. A model-free neural network-based classification system assigned each voxel a "nonmalignant" or "malignant" label based on the textural features. The classification results were compared via receiver operating characteristic (ROC) curve analysis with the manual lesion segmentation produced by two radiologists (observers 1 and 2). Results: The mean sensitivity and specificity of the classifier agreed with the mean observer 2 performance when compared with segmentations by observer 1 for a 95% confidence interval, using a two-sided t-test with α = 0.05. The results show that an area under the ROC curve (Az) of 0.99948, 0.99867, and 0.99957 can be achieved by comparing the classifier vs. observer 1, classifier vs. union of both observers, and classifier vs. intersection of both observers, respectively. Conclusion: This study shows that a neural network classifier based on 4D texture analysis inputs can achieve a performance comparable to that achieved by human observers, and that further research in this area is warranted. © 2007 Wiley-Liss, Inc.
  • Thumbnail Image
    ItemOpen Access
    Motion planning of a mechanical snake using neural networks
    (1998) Fidan, Barış
    In this thesis, an optimal strategy is developed to get a mechanical snake (a robot composed of a sequence of articulated links), which is located arbitrarily in an enclosed region, out of the region through a specified exit without violating certain constraints. This task is done in two stages: Finding an optimal path that can be tracked, and tracking the optimal path found. Each stage is implemented by a neural network. Neural network of the second stage is constructed by direct evaluation of the weights after designing an efficient structure. Two independent neural networks are designed to implement the first stage, one trained to implement an algorithm we have derived to generate minimal paths and the other trained using multi-stage neural network approach. For the second design, the intuitive multi-stage neural network back propagation approach in the literature is formalized.