Segmentation of Underwater Object in Videos

Video segmentation is a necessary step for object tracking. Existing methods that are used to extract object from the background based on an intensive sequence of searching all across the frames, thus this process performs lots of searching works result with low efficiency, whereas other methods obtain segmentation by clustering pixels which resulting in over-segmentation. Inspired by breakthroughs in semantic segmentation, in this paper, we propose to combine appearance and dynamic cues, which is a common conception and plays a key role in successfully segmenting objects in videos. To implement this idea, we combine Deep Convolutional Neural Network (DCNN) and optical flow information of two continuous frames. To overcome the difficulty of segmentation of underwater object in videos induced by the presence of different types of suspension particle from like the water droplets and dust particles to the poor lighting and over lighting conditions, In this work, Contrast-Limited Adaptive Histogram Equalization (CLAHE) and a simple color resign method are used to enhance details and reduce greenish and bluish effects. Some DCNN variants are applied to semantic segmentation and achieve great efficiency. Specifically, because DCNN can obtain different spatial scale information, as a DCNN variant, DeepLab gets a good performance in semantic segmentation. By using atrous convolution, DeepLab network's filters can observe greater receptive field without reducing the feature map dimension, therefore this structure keeps global and position information. Consequently, we compromises above mentioned methods, The optical flow estimation is carried out on the image processed by the CLAHE method, and the accurate segmentation results are obtained by using the DeepLab network. Experiments show good performance of our method.