Deep Learning based Classification Framework to save Right Whales

Abstract

Few decades ago, experienced Marine biologists started a campaign to safeguard Right Whales with a name of National Oceanic Atmospheric Administration. Different aerial campaigns were started by them, with the primary objective of studying their health and counting their population. Photographs were taken by helicopters and then these were compared with an online database. Manual comparing of Right Whales was very time consuming and lot of training and experience was needed to do it. So to overcome this issue NOAA in collaboration with Kaggle decided to launch a competition, the objective was to launch a system to monitor Right Whales and efforts must be done to free a Right Whale that has been accidentally caught in fishing gear. We only have 4544 training images in our dataset; training deep convolutional neural networks with such a small number is a real challenge. These photographs are taken by helicopters, so they are badly focused. They are taken at different times of day with different quality of cameras. Some images of Right Whales are of really bad contrast with poor exposure. Moreover the dataset was not balanced i.e. the number of pictures per whale varies a lot. We have about 20 whales with just 1 image, some whales have around 40 images, the average number of images per whale was around 10. With such a sparse distribution, it was really challenging for us to train our deep convolutional neural networks. To minimize the effect of small dataset, we have divided our problem wisely. Instead of training our neural network on whole images, which is no use to us, as most of the images contains ocean waves. We first localize head of the whale, in this way we can focus more on our desired feature, that is the callosity pattern on Whale’s head. After localizing head we find two points Bonnet and Blowhead, callosity pattern lies between these two points. We then aligns whale’s head in such a way, that Bonnet is on right blowhead is on left and whale’s head is pointing towards east. Now images are align, so our classifier can only focus on area of interest. This has improved our results significantly aswe have achieved an accuracy of 78.70%. The results can be improved if we have more images in the dataset or if we have more clearer images with better resolution and area of interest i.e Callosity Pattern more clearly visible.