Assessment of honey bee cells using deep learning

• Main Author: Alves, Thiago da Silva
• Publication Date: 2018
• Other Authors: Ventura, Paulo J.C., Neves, Cátia J., Candido Junior, Arnaldo, Paula Filho, Pedro L. de, Pinto, M. Alice, Rodrigues, Pedro João
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10198/18029
• topic: Deep learning; Artificial intelligence; Comb assessment

Temporal assessment of honey bee colony strength is required for different applications in many research projects. This task often requires counting the number of cells with brood and food reserves multiple times a year from images taken in the apiary. There are thousands of cells in each frame, which makes manual counting a time-consuming and tedious activity. Thus, the assessment of frames has been frequently been performed in the apiary in an approximate way by using methods such as the Liebefeld. The automation of this process using modern imaging processing techniques represents a major advance. The objective of this work was to develop a software capable of extracting each cell from frame images, classify its content and display the results to the researcher in a simple way. The cells’ contents display a high variation of patterns which added to light variation make their classification by software a challenging endeavor. To address this challenge, we used Deep Neural Networks (DNNs) for image processing. DNNs are known by achieving the state-of-art in many fields of study including image classification, because they can learn features that best describe the content being classified, such as the interior of frame cells. Our DNN model was trained with over 60,000 manually labeled images whose cells were classified into seven classes: egg, larvae, capped larvae, honey, nectar, pollen, and empty. Our contribution is an end-to-end software capable of doing automatic background removal, cell detection, and classification of its content based on an input image. With this software the researcher is able to achieve an average accuracy of 94% over all classes and get better results compared with approximation methods and previous techniques that used handmade features like color and texture.