When the number of horn flies that blood feed on cattle exceeds the economic threshold, they can adversely affect the health and wellbeing of their hosts. Excessive horn fly burdens also lead to reduced weight gain and, consequently, diminished profits for livestock producers. Effective management and treatment require reliable surveillance methods for estimating the degree of horn fly burden (i.e., counting the number of flies on cattle). Traditionally, these estimates are obtained through human visual estimation, either in-person or by counting images on a photo; however, these methods are costly both in terms of time and labor and are prone to subjectivity and bias. In contrast, automated methods can expedite the counting process and remove subjectivity and bias. To this end, a 2-stage method is presented here that uses computer vision and deep learning to identify the location of flies in digital images. The first stage segments the salient cow from all other parts of the image to remove flies on neighboring cattle from consideration. The second stage processes full-resolution patches of the original image and produces a heat map at the location of flies in the images. The method was trained on a set of 375 human-annotated images and tested on 120 images, where significant variation was observed amongst the human scorers. Counting results are compared to four separate human scorers and demonstrate that the neural network produces consistent results and that the method is reliable. Thus, the developed method can be used for monitoring changes in horn fly populations over time by anyone and allows for increased rigor and repeatability. An examination of individual images where the method was closest to and farthest from the human counts provides valuable insights regarding photographic processes that lead to success and failure.
Read full abstract