Peacock ocelli dataset creation and identification using dilated Convolutional Neural Networks
Recognition of peacocks in the wildlife is important for various reasons such as monitoring their behaviour, to know more about their habitat and to estimate their population. Some scientific research work suggests that peacock can be uniquely identified by its feather eyespots on the lower region of the feathers. The eyespots are also known as “ocelli”. For this, the process of counting the total number of feather eyespots and its location on the feather is required. Since peacock generally have more than a hundred eyespots, automating the task of feather eyespots counting is required. In this project an annotation tool is made which reduces the human effort in the counting and locating the eyespots on an image. Detection of eyespots is done using Circular Hough transformation. The primary use of the annotator is to make the dataset from images, as it stores the count of feather eyespots and the exact location of the eyespots. To make a continuous density map from the annotated points, adaptive Gaussian Kernel was used. This dataset of image patches and corresponding true density map was then studied on some Convolutional Neural Networks. Currently CSRNet- dilated Convolutional Neural Network is being studied on the dataset. The produced density map can then be used to localise the peacock eyespots.
Keywords: Hough Circle Transform, opencv.js, Gaussian Kernel, dilated Convolutional Neural Network