Retinal diseases are significant cause of visual impairment globally. In the worst case they may lead to severe vision loss or blindness. Accurate diagnosis is a key factor in the right treatment planning that can stop or slow the disease. The examination that can aid in the right diagnosis is Optical Coherence Tomography (OCT). OCT scans are susceptible to various noise effects which deteriorate their quality and as a result may impede the analysis of their content. In this paper, we propose a novel and effective method for OCT image denoising using a deep learning model trained on pairs of noisy and clean scans obtained by BM3D filtering. A comprehensive dataset of 21926 OCT scans, collected from 869 patients (1639 eyes), covering both healthy and pathological cases, was used for training and testing of the proposed scheme. The method was validated taking into account quantitative metrics concerning image quality. In addition, the proposed denoising scheme was evaluated by analyzing the impact of applying it in the eye disease classification based on Convolutional Neural Networks (CNNs) where we obtained the improvement of around 1-3 pp (percentage point). A separate dataset of 25697 scans collected from 1910 patients (2953 eyes) was used for this purpose. The conducted experiments have proved that the method can be applied as a preprocessing step in order to provide better disease classification results and can be useful in other OCT image analysis tasks. The proposed solution is much faster and perform better than the classical BM3D filter (over ninetyfold speed-up) and other related methods, especially when a big set of images needs to be processed at once. Furthermore, the use of the diverse dataset show the benefit over methods which are based on using only healthy scans for the training of the neural network.
The paper focuses on the application of new orthogonal transforms in digital watermarking. Novel types of transforms and their characteristics are presented. Potential methods for watermark embedding and recovery are also proposed. They assume embedding hidden information in the transform domains using the luminance channel of the original image. Image spectra are obtained by dividing the original image into smaller blocks that then are further processed by performing the forward transform operation. A watermark is embedded by modifying the spectral coefficients with relatively low values. Since there are various types of transforms, the latter process is realized in an adaptive manner. The proposed solutions were evaluated by measuring the level of visual distortion with respect to the total size of the inserted data. Additionally, the bit error rate (BER) in the recovery phase is also analyzed. The elaborated methods seem to be useful for applications in digital signal and image processing where high imperceptibility and low BER are of great importance. New orthogonal transforms were proved to be useful in watermarking tasks, and in some cases, they can even outperform the classic DCT approach.
Advanced Driver Assistance Systems (ADAS) are becoming more and more important in recent years. All relevant manufacturers equip their cars with various safety features and in some locations it is even mandatory to do so. One of such systems is Automatic Headlight Control (AHC) algorithm that prevents from accidental dazzling of other road users. Recognizing vehicle lights with high accuracy is crucial in this algorithm. Any improvement can result in the limitation of car accidents and dangerous situations on the road. The paper introduces a novel visual descriptor that describes previously binarized blob shapes of the analyzed objects. We also present our referenced AHC system that operates not only on the images but also on the set of other visual features. Adding new feature leads to the improvement of around 1.5 percentage point of vehicle lights classification accuracy (21% reduction of incorrectly classified light sources). It can also replace the already used features (four-time reduction from 32 to 8 feature values) and keep the classification accuracy at the similar level. The proposed solution has been tested with an extensive set of video sequences (total duration exceeding ten hours) under diverse weather conditions and from different types of roads. The proposed descriptor is shown to be far less computationally complex, and it is proved that it can outperform other state-of-the-art descriptors when analyzed individually.
The paper presents a novel data-embedding method based on the Periodic Haar Piecewise-Linear (PHL) transform. The theoretical background behind the PHL transform concept is introduced. The proposed watermarking method assumes embedding hidden information in the PHL transform domain using the luminance channel of the original image. The watermark is embedded by modifying the coefficients with relatively low values. The proposed method was verified based on the measurement of the visual quality of an image with a watermark with respect to the length of the embedded information. In addition, the bit error rate (BER) is also considered for different sizes of a watermark. Furthermore, a method for the detection of image manipulation is presented. The elaborated technique seems to be suitable for applications in digital signal and image processing where high imperceptibility and low BER are required, and information security is of high importance. In particular, this method can be applied in systems where the sensitive data is transmitted or stored and needs to be protected appropriately (e.g., in medical image processing).
In this paper, we propose a novel solution to the problem of insufficient training data in complex image classification tasks. The presented method efficiently recommends new learning samples with the use of contour-based descriptors and Scalable Vocabulary Tree (SVT) technique. The SVT is built from SIFT features and may be used for identification and localization of objects of interest in the analyzed query images. In conditions where the training data set is not large enough, the SVT's classification accuracy is satisfactory and comparable to that achieved by the dedicated CNN network. However, the SVT approach surpasses the classic convnet approaches due to its incremental learning capability. The classification certainty is further increased thanks to contour-based analysis using Point Distance Histogram (PDH) descriptor. At this stage, the PDH representation of the analyzed object is matched with the information stored in the pattern database. This database contains a set of 360 shape outlines for each object that are extracted from its 3D model. This leads to higher classification accuracy of the presented approach.