Traffic sign recognition techniques aim to reduce the probability of traffic accidents by increasing road and vehicle safety. These systems play an essential role in the development of autonomous vehicles. Autonomous driving is a popular field that is seeing more and more growth. In this study, a new high-performance and robust deep convolutional neural network model is proposed for traffic sign recognition. The stacking ensemble model is presented by combining the trained models by applying improvement methods on the input images. For this, first of all, by performing preprocessing on the data set, more accurate recognition was achieved by preventing adverse weather conditions and shooting errors. In addition, data augmentation was applied to increase the images in the data set due to the uneven distribution of the number of images belonging to the classes. During the model training, the learning rate was adjusted to prevent overfitting. Then, a new stacking ensemble model was created by combining the models trained with the input images that were subjected to different preprocessing. This ensemble model obtained 99.75% test accuracy on the German Traffic Sign Recognition Benchmark (GTSRB) dataset. When compared with other studies in this field in the literature, it is seen that recognition is performed with higher accuracy than these studies. Additively different approaches have been applied for model evaluation. Grad-CAM (gradient weighted class activation mapping) was used to make the model explainable. Evidential deep learning approach was applied to measure the uncertainty in classification. Results for safe monitoring are also shared with SafeML-II, which is based on measuring statistical distances. In addition to these, the migration test is applied with BTSC (Belgium Traffic Sign Classification) dataset to test the robustness of the model. With the transfer learning method of the models trained with GTSRB, the parameter weights in the feature extraction stage are preserved, and the training is carried out for the classification stage. Accordingly, with the stacking ensemble model obtained by combining the models trained with transfer learning, a high accuracy of 99.33% is achieved on the BTSC dataset. While the number of parameters the single model is 7.15 M, the number of parameters of the stacking ensemble model with additional layers is 14.34 M. However, the parameters of the models trained on a single preprocessed dataset were not trained, and transfer learning was performed. Thus, the number of trainable parameters in the ensemble model is only 39,643.

Segmentation of vasculature in retinal fundus image by level set methods employing classical edge detection methodologies is a tedious task. In this study, a revised level set-based retinal vasculature segmentation approach is proposed. During preprocessing, intensity inhomogeneity on the green channel of input image is corrected by utilizing all image channels, generating more efficient results compared to methods utilizing only one (green) channel. A structure-based level set method employing a modified phase map is introduced to obtain accurate skeletonization and segmentation of the retinal vasculature. The seed points around vessels are selected and the level sets are initialized automatically. Furthermore, the proposed method introduces an improved zero-level contour regularization term which is more appropriate than the ones introduced by other methods for vasculature structures. We conducted the experiments on our own dataset, as well as two publicly available datasets. The results show that the proposed method segments retinal vessels accurately and its performance is comparable to state-of-the-art supervised/unsupervised segmentation techniques.

Machine learning techniques, especially deep learning, have made great achievements in computer vision including image denoising recently. However, in most convolutional neural network (CNN) based methods presented for image denoising, convolutional kernels are considered for only one scale and more scales are neglected mostly. Studies on multilevel feature extraction treat these features as if they have the same importance and do not use a mechanism such as feature attention for their weighting. Also, for effective noise removal, edge information is used as prior knowledge, but texture information is generally disregarded. This study has focused on these shortcomings and introduced a new attention-based CNN for image denoising. The main contributions of this study are as follows: First, we propose a CNN-based network to extract Local Binary Pattern (LBP) from the noisy image for texture information. So, we use texture information as prior knowledge for the preservation of details in the evolved image during the denoising process. Besides we propose a new multilevel feature extraction block to get different level features. After extracting multilevel features using feature attention, we weight these different levels of features. In addition to this, we introduce a multilevel feature attention network (MLFAN) for noise removal by combining them. The comprehensive experimental results show that our MLFAN noise reduction network can effectively remove Gaussian noise from images and compared with some state-of-the-art denoising methods, it outperforms in terms of both quantitative and qualitative evaluations. For Set12 grey image set, and McMaster color image set, MLFAN gives PSNR={33.08, 30.75, 27.56}, SSIM={0.9087, 0.8702, 0.7939} and PSNR={35.08, 32.68, 29.47}, SSIM={0.9288, 0.8956, 0.8263} respectively for noise level σ = {15, 25, 50}.

Degradations frequently occur in archive films that symbolize the historical and cultural heritage of a nation. In this study, the problem of detection blotches commonly encountered in archive films is handled. Here, a block-based blotch detection method is proposed based on a visual saliency map. The visual saliency map reveals prominent areas in an input frame and thus enables more accurate results in the blotch detection. A simple and effective visual saliency map method is taken into consideration in order to reduce computational complexity for the detection phase. After the visual saliency maps of the given frames are obtained, blotch regions are estimated by considered spatiotemporal patches—without the requirement for motion estimation—around the saliency pixels, which are subjected to a prethresholding process. Experimental results show that the proposed block-based blotch detection method provides a significant advantage with reducing false alarm rates over HOG feature (Yous and Serir, 2017), LBP feature (Yous and Serir, 2017), and regions-matching (Yous and Serir, 2016) methods presented in recent years.

In this article, an improved approach for editing color images has been presented. In this approach, using the Poisson equation, a guided vector field is created by employing source and target images within a selected region at the first step. Next, the guided vector is used in generating the result image. Most of the existing techniques in the literature perform image editing without taking color information into account. However, without utilizing color information, image geometry cannot be created properly in some cases which may result in unsatisfactory results. Unlike the existing methods, the proposed study utilizes all the information contained in each of the color channels in computing gradient norm and performing inpainting process. The test results show that the suggested technique generates satisfactory results in editing color images.

Traffic sign recognition has been one of the indispensable issues of Advanced Driver Assistance Systems. In this study, a new CNN model for traffic sign recognition based on deep learning is proposed. The proposed model has low number of parameter and high accuracy compared to most studies in the literature. Initially, in preprocessing stage, different color spaces are tried for the input image, and their combinations are given to the network together. Color spaces used in the study are RGB, CIELab, RIQ and LGI. In addition, the accuracy results were compared by experimenting on the input image dimensions. Additionally, data augmentation was applied during the training phase. As a result, 98.84% accuracy was obtained by giving the input image with RIQ and LGI color space to the network. The number of parameters is 0.95 M.