Machine learning - neural network-based algorithm
This method was applied using the Fiji platform, the DenoiSeg tool from the CSBDeep plug-in, and the neural network algorithm for instance segmentation (Buchholz et al. 2020). The machine learning process requires an appropriate graphics card (e.g. NVIDIA) and the installation of several drivers and software that operate in the background of the main software (CUDA Toolkit, GPU support, TensorFlow, cuDNN SDK, Phyton etc.); they differ depending on the computer’s firmware and operating system. After preparing the computer we need to manually prepare labelling images in the graphics software - photos enabling objects be sufficiently visible against the background. Then we need to pair them with the raw photos. To learn this neural network, a small number (2 - 10) of training data are needed. The most time-consuming aspects are the image preparation (training data) and the neural network learning process. Depending on the power of the computer, the latter may take a few, twelve or even more hours. This process produces a model that can be used for prediction (Buchholz et al. 2020; Schroeder et al. 2020). After running the model in Fiji software, cleaning the photo should take a few seconds, and then it should take only a few more seconds to count the birds using the Analyze Particles command in Fiji (Sandhya et al. 2011). The step-by-step procedure is described on the website devoted to this method, from which one can download sample data and perform the machine learning process on one’s own computer: https://imagej.net/DenoiSeg