Implementation and application of the wavelet transform into deep learning techniques for texture classification and object detection

Abstract

Texture characterization in digital images has become an analysis tool in computer vision. Texture in visual perception is a very important physical property since it provides information about the structural composition of surfaces and objects in the image. This research involves two areas of knowledge, wavelet analysis and deep learning, both of which functioned as feature extraction methods for image processing with textures and materials. This work aimed to study the adaptability of deep learning with wavelet analysis and implement a detection and classification system for aerial navigation. The first approach analyzes the extracted information (spatial domain vs. wavelet domain) in object detection and aerial navigation. In addition, to evaluate the learning performance of a binary classifier. In the second approach, a multi-class classifier is proposed for the following databases: KTH-TIPS-2B (KT2B), Describable Textures Dataset (DTD), and Flickr Material Database (FMD). The possibility of merging both domains is evaluated since Convolutional Neural Networks (CNNs) do not learn spectral information, important information for texture recognition. In the third approach, a classification system for textured objects in aerial navigation tasks is implemented, where texture is involved as a physical property of the object. A classification model is developed using the knowledge transfer method and wavelet features. In the fourth approach, it is shown that internal pooling layers often lead to information loss. A classification system with a new pooling method called Discrete Wavelet Transform Pooling (DWTP) is proposed to solve this problem. The combination of these methods achieves acceptable classification performance. The learning plots reflect that all three datasets show learning generalization. In addition, the images obtained from the virtual environment show learning generalization for some classes in the DTD database. Moreover, the fusion of deep learning with wavelet analysis is recommended for small datasets of images with textures. Due to the limitation of learning about spectral information that is lost in conventional CNNs. Furthermore, it is argued that this helps to eliminate overfitting. The results show that it is possible to integrate this methodology into the technological development of applications, such as image classification or restoration tasks and object detection.