Unsupervised Image Segmentation model based on W Net architecture and Conditional Random Field for Underwater Images
Esra Sorkty Saleh, T.P. Mithun Haridas, M. H. Supriya · 2021
Under water imaging has been extensively used in monitoring the marine ecosystems for quite some time. The automated learning systems that are much explored in marine ecosystem are mainly supervised methods. Supervised methods even though require less human intervention, still demands considerable manual effort for the careful design of the dataset. The large amount of labelled and annotated data needed for proper learning, as well as the attention required for designing the ground truth dataset for efficient training, makes the supervised learning approach less attractive, especially in the underwater scenarios. This work explores the possibility of unsupervised learning methods for the analysis of the underwater images using segmentation algorithms, which does not need ground truth annotations and is much more scalable. Images and videos from Fish4Knowledge database is utilized for this work to train the unsupervised architecture based on W-Net architecture. The autoencoder architecture is analyzed considering the segmented output from the first half and reconstructed image from the second half using soft N-cut loss and reconstruction loss respectively. Post processing method of Conditional Random Field is also applied for smoothing the initial segmentation. The performance is analyzed using the learning curves and the metrics such as VI, PRI and SC.