Texture Recognition Using InceptionNeXt-based Texture Perception Network
Tapan Kumar Sahoo, Gaurav Mawari · 2024
Recognizing textures remains challenging in images due to variations in fundamental elements (primates) and characteristics (attributes) caused by spatial context. Existing Convolutional Neural Network (CNN) approaches typically use detailed local descriptions with unordered aggregation for spatial layout inconsistency. However, these methods overlook the inherent structure relationships between primates and the semantic meanings conveyed by attributes, which are crucial for texture. This paper suggests a newfangled approach for texture cognition using Inception NeXt, a deep learning architecture known for its long-range modelling capabilities. Inception NeXt uses large-kernel convolutions to capture broader contextual information within textures. We address the computational efficiency boundaries associated with large kernels by incorporating an Inception-inspired decomposition strategy. This approach decomposes the large kernels into smaller, parallel branches, enhancing efficiency while maintaining performance. Our model extracts features by analyzing the relationships between these primates and attributes within the spatial context of the texture. The Inception NeXt architecture facilitates the capture of long-range dependencies between texture elements, leading to more robust feature representations. Experiments on two challenging texture databases demonstrate the effectiveness of our approach, achieving superior performance in terms of accuracy, robustness, and efficiency compared to traditional CNN-based methods.