Harnessing the Feature Vectors of VGG-16 for Birds Detection under Geometric Distortions and Brightness

The geometric distortions and brightness shifts (GDB-effects)-due to the interference of the movements of the birds and the environmental light-are unavoidable when the birds' activities are video-recorded in the nature. In this paper, VGG-16 neural network is adapted to extract features from the images of backyard birds-cardinals and sparrows-that are GDB-affected. An augmented dataset that consists of GDB-affected images of cardinals and sparrows is also presented. An artificial neural network (ANN) that is trained on the VGG-16 features of this dataset has been studied. A simulation study shows that the VGG-16 features are robust and not affected by GDB-effects, and highly efficient for training an ANN model to classify cardinals and sparrows. It also reveals that the feature vectors of VGG-16 satisfy the Gaussian Mixture Model. Hence, the proposed modeling and simulation delivers a robust ANN model that classifies and detects cardinals and sparrows in the nature.