Akademik Veri Yönetim Sistemi
KNOWLEDGE-BASED SYSTEMS, cilt.327, 2025 (SCI-Expanded, Scopus)
Since camouflaged objects in camouflage images have weak boundaries and very close texture, color and pattern characteristics with the background, their detection and classification are very challenging problems. Traditional methods, which are widely used in the literature, are insufficient to solve this problem. Therefore, pre-trained transfer learning (TL) architectures were used to calculate classification performances. In this study we propose a novel dual-branch transfer learning architecture that integrates DenseNet201 and MobileNet models via a concatenation-based feature fusion strategy. This design allows the model to leverage multi-level semantic features from both networks, enhancing its ability to distinguish camouflaged objects in complex scenes. In deep learning-based architectures, it is very important to have enough data for classification success. However, this good performance is usually based on a sufficient amount of data. Insufficient data can lead to low classification performance or problems such as overfitting. Therefore, the number of raw training images was increased by applying data augmentation to the training images in the COD10K camouflage dataset used in this study. Additionally, augmented training data provided by ERVA 1.0, a challenging camouflage dataset, was used with ERVA 1.0 test data. As a result of the experimental studies, the DenseNet201 model showed the best classification performance with an accuracy of 97.67% for the classification task on the COD10K dataset and the DenseNet201 model showed the best classification performance with an accuracy of 98.49% on the ERVA 1.0 dataset. The study also combines pre-trained TL architectures with different combinations to create a new concatenation-based pre-trained approach with more extensive feature extraction and generalization capabilities. In this context, the architectures that perform the best classification performance are combined in different combinations. With these combinations, the best results were obtained with 98.41% accuracy for the COD10K and 98.83% accuracy for the ERVA 1.0 with the concatenation-based DenseNet201 + MobileNet model.