Comparative study on classification of textures using lightweight deep learning architectures based on transfer learning technique

Texture classification proposed in this paper develops deep learning (DL) models with Transfer learning (TL) and fine-tuning method on a cloud-based platform. In this paper, the lightweight deep learning architectures are compared and investigated between MobileNetV2 and NASNetMobile based on transfer learning technique. The standard texture database KTH-TIPS2b, and Kylberg Texture database v1.0 are used for classification. These datasets are split up into three different random ratios as training, validation, and testing. For each split ratio, two different Mobile Deep Learning (MDL) architectures are evaluated, whose weights are trainable from ImageNet. The observations infer the best-split ratio based on performance among the pre-trained models. These models are classified using SoftMax classifier. An Adam Optimization algorithm was utilized to enhance the model’s performance and to accelerate the learning speed. As the experiments are conducted with gray and color texture databases, the findings demonstrate 97% accuracy with the best-split ratio that greatly reduces memory usage and computation.