Local Intensity Gradient Directional Pattern Based Coral Image Classification using Deep Neural Network

Coral reefs are one of the world's most biologically diversified ecosystems on the earth that supports most of the marine life. The identification of coral species is essential in the conservation and monitoring process. Hence, an accurate automatic coral classification system is necessary to identify different types of corals. In this work, a new feature extraction method Local Intensity Gradient Directional Pattern (LIGDP) is proposed to extract features from coral images. The proposed method extracts the two maximum intensity changes for each pixel in the image by mapping four directional orientation approach i.e for each 3ⅹ3 intensity value two more 3x3 intensity values are extracted from four directional values. Finally, the LIGDBP pattern integrates the three 3x3 local grid pixel values extracted. To classify the coral images a new deep convolution neural network (CNN) is proposed with five layers. This custom neural network increases the classification accuracy with less number of parameters. The experiments were conducted with various coral image and video datasets to show the efficiency of the proposed technique. Moreover, the LIGDBP feature descriptor with a custom CNN model is compared with the existing techniques to demonstrate the efficiency of the proposed technique. The proposed model achieves a higher classification rate of 100% on EILAT-2 coral dataset.