Bees detection on images: Study of different color models for neural networks

Abstract

This paper presents an approach to bee detection in video streams using a neural network classifier. We describe the motivation for our research and the methodology of data acquisition. The main contribution to this work is a comparison of different color models used as an input format for a feedforward convolutional architecture applied to bee detection. The detection process has is based on a neural binary classifier that classifies ROI windows in frames taken from video streams to determine whether or not the window contains bees. Due to the type of application, we tested two methods of partitioning data into training and test subsets: video-based (some video for training, the rest for testing) and individual based (some bees for training, the rest for testing). The tournament-based algorithm was implemented to aggregate the results of classification. The manually tagged datasets we used for our experiments have been made publicly available. Based on our analysis of the results, we drew conclusions that the best color models are RGB and 3-channeled color models: RGB and HSV are significantly better than black & white or the H channel from HSV.