Fine-grained textural detail enhancement: concatenating convolutional neural network features with adaptive fuzzy logic

Research based on textural analysis has gained substantial consideration due to its significance in classification tasks. Hence, we introduce a Textural descriptor that derives differential textural features based on application requirements to enhance the discriminative nature of classification tasks. This research work leverages the benefits of pre-trained models including Google Net, Alex Net, ResNet, and Inspection ResNet to perform textural feature extraction and our approach concatenates the features obtained by the CNN architecture for acquiring fine-grained textural details of the textural images. Furthermore, the fuzzy logic-based adaptive feature compaction approach is incorporated to adjust the feature compaction level in adapting to the changing application requirements depending upon the variables including feature extraction time, error margin, and granularity. Also, the integration of attention vectors by generating region masks and assigning important weights based on changing needs identifies specific region of interest in the images, facilitating robust textural feature extraction. The robust textural features extracted by the Textural descriptor are applied to the SVM classifier to classify the textural features for performance evaluation. Here, we employed three textural image datasets namely, the KTH-TIPS dataset, the OUTEX_TC_00013-l dataset, and the Kylberg dataset, and utilized different previous textural analysis approaches for comparison. Our approach obtains 98.25% accuracy, 97.8% precision, and 95.83% recall which is evidence that the texture descriptor by capturing remarkable textural features enhanced the discriminative ability of the SVM classifier. Additionally, the dynamic nature of our approach not only surpassed the conventional techniques but also provided significant advancement in the domain of textural analysis.