Detection and recognition camouflaged objects in images: The case of butterfly

Camouflaged objects in camouflage images are difficult to detect because they have similar texture, pattern and colour characteristics to the background. Since they have weak boundaries and patterns similar to the background, existing binary segmentation solutions cannot easily cope with the problem of camouflaged object detection. Camouflaged object detection (COD) aims to detect objects with a high degree of similarity to the background. In this thesis, ERVA 1.0 original camouflage dataset is created and used. In this thesis, a two-stage solution is presented for the COD problem: segmentation and object recognition. The texture features of all test images on the ERVA 1.0 dataset are extracted using Gabor Filter for segmentation. These extracted features are clustered with the K-means algorithm and the original image is divided into different regions according to the texture features. LBP and Euclidean distance calculation were used to detect the objects in these regions. Then, pretrained models from deep learning techniques were used to predict the type of the object. In the study, the segmentation success rate was 87.89% with the Structural Similarity method and 83.64% with the Dice Similarity Coefficient method. Deep learning pretrained models were used to determine the object type obtained after segmentation. Experiment 1 was performed with un augmented data and Experiment 2 was performed with augmented data by applying data augmentation method. The highest success rate for Experiment 1 was 92.29% with the InceptionResNetV2 model and the highest success rate for Experiment 2 was 94.81% with the DenseNet121 model.