Automated Texture-Based Recognition of Corals in Natural Scene Image

Current coral reef health monitoring efforts rely on biodiversity data. Although cutting-edge imaging technology enables reliable and automatic collection of such data, simple RGB digital photography in combination with manual image annotation remains a popular solution. Unlike other acquisition methods, close range visible light imaging yields detailed species and surface coverage data, and requires cheaper equipment. Moreover, images acquired in the last few decades are limited to mere RGB photographs or analog VHS video, and contain important data for long term analysis. Unfortunately, manual expert labeling has become problematic due to the high volume of images and the lack of human resources available. Consequently, coral reef biodiversity data currently available is based mostly on small sample analysis. Previous automatic benthic image annotation systems have yielded unsatisfactory results compared to human performance for the same task. This is partly due to the high diversity of complex textures found in these images. We hypothesize that these complex textures require different features to be properly characterize. Motivated by the need for an improved automated benthic image annotation system, this work proposes a new approach based on a combination of multiple state-of-the art texture recognition methods. Firstly, methods to correct and enhance images will be investigated. Secondly, various state-of-the-art texture features will be used to overcome the texture diversity challenge: many statistical features, local binary patterns, textons, vector-quantized Scale-Invariant Feature Transform (SIFT) using the Improved Fisher Vector (IFV) method, Deep Convolutional Activation Feature (DeCAF), amongst others. Thirdly, a multi-classifier fusion method is proposed to efficiently aggregate the information from these multiple texture representations using a score-level fusion. Fourthly, rejection will be applied to further enhance accuracy. The results on the AIMS dataset (Australian Institute of Marine Science) and MLC2008 (Moorea Labeled Corals 2008) containing respectively 75 825 and 131 260 coral texture patches show that the proposed multi-classifier fusion method outperforms any other single method for the task of benthic image labeling.