Abstract: Research in texture recognition often concentrates on recognizing textures with intraclass variations such as illumination, rotation, viewpoint and small scale changes. In contrast, in real-world applications a change in scale can have a dramatic impact on texture appearance, to the point of changing completely from one texture category to another. As a result, texture variations due to changes in scale are amongst the hardest to handle. In this work we conduct the first study of classifying textures with extreme variations in scale. To address this issue, we first propose and then reduce scale proposals on the basis of dominant texture patterns. Motivated by the challenges posed by this problem, we propose a new GANet network where we use a Genetic Algorithm to change the units in the hidden layers during network training, in order to promote the learning of more informative semantic texture patterns. Finally, we adopt a FVCNN (Fisher Vector pooling of a Convolutional Neural Network filter bank) feature encoder for global texture representation. Because extreme scale variations are not necessarily present in most standard texture databases, to support the proposed extreme-scale aspects of texture understanding we are developing a new dataset, the Extreme Scale Variation Textures (ESVaT), to test the performance of our framework. It is demonstrated that the proposed framework significantly outperforms gold-standard texture features by more than 10% on ESVaT. We also test the performance of our proposed approach on the KTHTIPS2b and OS datasets and a further dataset synthetically derived from Forrest, showing superior performance compared to the state of the art.
Introduction: The "Distance Problem" in Texture Recognition
In computer vision, texture recognition is a foundational task, but it faces a significant challenge when dealing with distance. Most existing algorithms are trained on datasets with modest scale variations. However, in real-world scenarios—such as autonomous drone navigation or long-range surveillance—a texture’s appearance can change so drastically that it effectively transitions into a different category. For instance, a forest viewed from miles away resembles a green carpet, while up close, it is a complex collection of bark and leaves.
The authors identify two major gaps in current research: standard Convolutional Neural Networks (CNNs) are not naturally robust to scale changes of two or more orders of magnitude, and existing benchmarks (like Brodatz or CUReT) lack the extreme scale variations found in the wild.
The GANet Framework
To address these challenges, the researchers proposed a multi-stage pipeline centered around a novel architecture called GANet (Genetic Algorithm Network). Instead of relying solely on standard backpropagation, GANet uses a Genetic Algorithm (GA) to optimize the internal architecture of a VGG-19 based network during training.
Genetic Optimization
The goal of this GA-based approach is to increase the number of "semantic" units—filters that respond to meaningful patterns—while suppressing noisy or incomplete "non-semantic" filters. The system trains three parallel CNNs, applying Crossover (swapping filter strings between networks) and Mutation (replacing non-semantic filters with semantic ones from a different net). This acts as a targeted form of dropout that forces the network to learn more robust, informative features.
Scale Proposal and Reduction
To handle scale invariance, the system employs a scale-proposal mechanism. It uses the BING algorithm to identify "texels" (texture elements) across an image pyramid. Since an image might generate thousands of scale candidates, the authors reduce these to a few "dominant" scales by clustering similar texels. This ensures the classifier focuses on the most representative zoom levels, significantly improving computational efficiency.
Global Representation with FV-CNN
Finally, the features evolved by GANet are pooled using Fisher Vectors (FV). This creates a global texture representation that is orderless, making it more flexible than standard fully connected layers for describing repeating patterns across various scales.
Experimental Results and Datasets
The performance of GANet was evaluated on several datasets, including the newly introduced ESVaT (Extreme Scale Variation Textures) dataset, which contains over 15,000 images specifically designed to test extreme scale changes.
Key results include:
- ESVaT Performance: GANet achieved 67.9% accuracy, significantly outperforming standard FV-CNN (56.2%) and traditional descriptors like LBP (40.5%).
- SForrest Benchmark: On a synthetic dataset derived from Forrest, the method reached 96.3%, a massive lead over the next best state-of-the-art method (85.3%).
- Generalization: Even on standard datasets with smaller scale variations like KTHTIPS2b, GANet showed a 4-5% improvement, proving that GA-based filter evolution produces higher-quality feature extractors regardless of the scale range.
Conclusion
GANet demonstrates that incorporating genetic algorithms into neural network training can effectively promote the learning of semantic patterns necessary for robust texture classification. By combining this with a dominant-texel scale proposal mechanism, the framework overcomes the traditional "distance problem" in computer vision. Future work will focus on refining "scale boundary" detection to help autonomous systems better understand when a material's visual identity changes due to distance.
Summary updated