3.2 Using Convolutional Neural Nets to filter out
unsuitable images
The optimal conditions for the Convolutional Neural Net (CNN) trained to
recognise wild dogs standing up were two convolutional layers, with
kernel sizes of 32 and 64, respectively, and a learning rate of
10-5. This model achieved a training accuracy of
100%, a validation accuracy of 91% (95% C.I. 90 – 92), and a testing
accuracy of 90% (95% C.I. 88 – 91, Table 2). For the CNN designed to
separate images of the left and right flanks, the optimal conditions
were three convolutional layers, one with a kernel size of 64 and two
with kernel sizes of 32, with a learning rate of 10-4.
Its training, validation and testing accuracy were 100%, 96% (95%
C.I. 95 - 97), and 95% (95% C.I. 94 - 96), respectively.