Texture-based instance retrieval is typically performed on images that present a single texture pattern and is mainly applied to the retrieval of fabrics or textiles. In this work, we apply it to indoor scene images that typically present many different texture patterns, which constitutes a more challenging problem. Such retrieval systems, together with the retrieval of faces and objects, can be used as a valuable tool for evidence matching in crime scene investigation. Even though recent deep learning-based approaches have made significant improvement in many computer vision tasks, texture retrieval remains an open problem. In this work, we introduce a Fourier-based approach, in which spatial images and their discrete Fourier transform maps are combined to derive a novel texture representation. We further present a new and efficient texture-based image retrieval framework based on region proposal networks, convolutional autoencoders and transfer learning, in which we extract the features from the latent space layer of the encoder as texture descriptors. The experimental results on four datasets: TextileTube, Outex, USPtex and Stex, validated the effectiveness of our proposed method, yielding better results than the current state of the art.

基于纹理的实例检索通常在呈现单个纹理图案的图像上执行，并且主要应用于织物或纺织品的检索。在这项工作中，我们将其应用于通常呈现许多不同纹理图案的室内场景图像，这构成了一个更具挑战性的问题。这种检索系统以及面部和物体的检索，可以用作犯罪现场调查中证据匹配的有价值的工具。即使最近基于深度学习的方法在许多计算机视觉任务中取得了显着改善，但纹理检索仍然是一个未解决的问题。在这项工作中，我们介绍了一种基于傅立叶的方法，其中将空间图像及其离散的傅立叶变换图进行组合，以得出新颖的纹理表示。我们进一步提出了一种新的高效的基于纹理的图像检索框架，该框架基于区域提议网络，卷积自动编码器和传递学习，其中我们从编码器的潜在空间层提取特征作为纹理描述符。在四个数据集（TextileTube，Outex，USPtex和Stex）上的实验结果验证了我们提出的方法的有效性，比当前的最新技术产生了更好的结果。