Convolutional Neural Networks (CNNs) have recently been proposed as a solution in texture and material classification in computer vision. However, inside CNNs, the internal layers of pooling often cause a loss of information and, therefore, is detrimental to learning the architecture. Moreover, when considering images with repetitive and essential patterns, the loss of this information affects the performance of subsequent stages, such as feature extraction and analysis. In this paper, to solve this problem, we propose a classification system with a new pooling method called Discrete Wavelet Transform Pooling (DWTP). This method is based on the image decomposition into sub-bands, in which the first level sub-band is considered as its output. The objective is to obtain approximation and detail information. As a result, this information can be concatenated in different combinations. In addition, wavelet pooling uses wavelets to reduce the size of the feature map. Combining these methods provides acceptable classification performance for three databases (CIFAR-10, DTD, and FMD). We argue that this helps eliminate overfitting and that the learning graphs reflect that the datasets show learning generalization. Therefore, our results indicate that our method based on wavelet analysis is feasible for texture and material classification. Moreover, in some cases, it outperforms traditional methods.
22 December 2020 in Programming and Computer Software
This paper proposes a first approach based on wavelet analysis inside image processing for object detection with a repetitive pattern and binary classification in the image plane, in particular for navigation in simulated environments. To date, it has become common to use algorithms based on convolutional neural networks (CNNs) to process images obtained from the on-board camera of unmanned aerial vehicles (UAVs) in the spatial domain, being useful in detection and classification tasks. CNN architecture can receive images without pre-processing, as input in the training stage. This advantage allows us to extract the characteristic features of the image. Nevertheless, in this work, we argue that characteristics at different frequencies, low and high, also affect the performance of CNN during training. Thus, we propose a CNN architecture complemented by the 2D discrete wavelet transform, which is a feature extraction method. Wavelet analysis allows us to use time-frequency information through a multiresolution analysis. Therefore, we have investigated the combination of multiresolution analysis, via wavelets, in conjunction with CNN architectures, to use the information in the wavelet domain as input to the training stage. The information improves the learning capacity, eliminates the overfitting, and achieves a better efficiency in the detection of a target. Also, our learning model was evaluated in the aerial navigation of a drone.
1 January 2021 in Proceedings of the Institute for System Programming of the RAS
Предлагается первый подход, основанный на применении вейвлет-анализа при обработке изображений с целью обнаружения объектов c повторяющимися чертами и двоичной классификации в плоскости изображения, в частности, для навигации в симулируемых средах. На сегодняшний день стало привычным использовать алгоритмы на основе сверточных нейронных сетей (Convolutional Neural Networks, CNN) для обработки изображений, полученных с бортовых камер беспилотных летательных аппаратов (Unmanned Aerial Vehicles, UAV), в пространственной области, что способствует решению задач обнаружения и классификации. Архитектура CNN позволяет обучать сеть, используя в качестве входных данных изображения без предварительной обработки. Это позволяет извлекать характерные признаки изображения. Тем не менее, в этой работе мы утверждаем, что спектральные характеристики изображений на разных частотах, низких и высоких, также влияют на производительность CNN во время обучения. Мы предлагаем архитектуру CNN, дополненную двумерным дискретным вейвлет-преобразованием как методом выделения признаков. Такая информация улучшает способность сети к обучению, устраняет переобучение и обеспечивает более высокую эффективность при обнаружении цели.
An essential part of aerial navigation is to detect and locate targets of interest. This research chapter focuses on object classification, where the important feature is texture. We present a classification model that uses transfer learning and wavelet-based features as an additional feature extraction method. This model is trained with the Describable Texture Database (DTD), a 53% accuracy is obtained. We developed a virtual world in the Gazebo simulator to validate the results. The virtual world allowed us to create a new Synthetic Aerial Dataset of Textured Objects (SADTO) and integrate the Parrot AR.Drone model with its flight features. ROS Kinetic is used to send information from the on-board camera of the drone (virtual environment) to the PC. The images from the environment show a generalization of the knowledge for some classes of the database. <!----> <!----> <!----> <!---->
While convolutional neural networks (CNNs) excel at image classification, their generalization across domains and robustness to nonlinear degradation remain challenges in art media classification (AMC). To address these challenges, this article presents a dual-stage analytical framework: first, an evaluation of seven discrete CNN architectures—ranging from VGG16 to ConvNeXt—subjected to domain shift using the New Spain (Mexico) Art Media Dataset; and second, a formal robustness analysis using an artistic corruption benchmark (Art-C). This benchmark simulates nonlinear degradations, including cracking, oxidized varnish, and pictorial abstraction. Our results demonstrate that while deep convolutional representations maintain acceptable transferability (accuracy ), significant variability exists in architectural stability (mean 0.0607) under progressive stochastic degradation. Notably, Xception exhibited the highest robustness (Art-mCE = 0.8039), whereas VGG16 showed the greatest relative performance decay. Severity analysis further indicates that structural perturbations induce higher error rates than chromatic shifts, suggesting that CNNs are more sensitive to topological features (depth and residual connections) than color-space distributions. We provide quantitative evidence characterizing the relationship between architectural topology and empirical stability in non-natural image domains.