Dual-stream deep learning framework integrating texture and shape features for robust image classification

Abstract

Image classification is a vital component in applications like preventing infestations, medical diagnosis, handwriting recognition and many more. Many present methods only use shape or texture features, which makes them less effective in the real world. Texture-based methods capture a lot of surface characteristics, but they are affected by changes in lighting and background noise. Shape-based methods, on the other hand, are better at dealing with noise but sometimes lose important texture information. This research proposes a single dual-stream deep learning architecture that learns texture and shape representations parallelly. A ResNet-50 network extracts texture features from raw images, and a parallel ResNet-50 network processes skeletonized images to get shape characteristics. The extracted texture and shape features are combined through weighted feature fusion to capture complementary information for pest classification. To improve class separability and generalization, a Support Vector Machine (SVM) is employed to classify the features obtained from both streams. The system was tested on a dataset of pest images that included both primary and secondary dataset. The proposed framework achieved a classification accuracy of 94.0% on the train-test evaluation. To obtain a more reliable estimate of model generalization, five-fold stratified cross-validation was also performed, yielding a mean accuracy of 92.84% ± 0.97%. The validation of the novel approach was carried out using random picked images from the primary dataset. The suggested method combines texture and shape information into one architecture, making it a strong and adaptable solution for a wide range of image identification challenges.