Abstract:
Image recognition with deep learning focuses on identifying one or more specific
objects, class or feature in a specified image or video frame. Deep learning has received
a lot of attention recently and is yielding unprecedented results. Image recognition is used
in different applications, such as Medical diagnosis, Automatic inspection of
manufacturing products and tasks like pedestrian detection. In the dairy industry, teat
spraying is essential in order to prevent the spread of diseases like mastitis. However, the
current automatic teat spraying/cup attachment methods have a lot of issues. For example,
teat location coordinates are assumed to be fixed and are pre-stored while in reality, it is
more likely that the cow might be more mobile, hence it is a flawed approach. Secondly,
IR based systems are inconsistent due to the light absorption property of IR which has no
guidance for exact detection. To address this problem, it is required to attach the cup or
spray the teats only when the udder is detected. This thesis makes use of deep learning
for the detection of cow udder which can be of great importance in the dairy industry
which is currently relying on conventional methods that are inefficient and inconsistent.
Though Deep Convolutional Neural Network is the most advanced technique for image
recognition but it demands a great deal of training time and computational power.
Focusing on the problem of image recognition with limited time for training and restricted
computing power in mind, a technique called “Transfer Learning” is used in this thesis.
This technique makes use of the pre-trained model in order to accelerate the process of
learning. Trainable parameters from Xception Deep Convolution Neural Network
(DCNN) model were transferred for identifying cow udder. The model training was done
in three stages; the data collection or dataset building stage which includes acquiring cow
udder images and manual segmentation then training stage and finally the
validation/testing phase. The model is trained on our custom made cow udder dataset.
With overall miou_1.0 of 0.88. The model was successful in accurately detecting udder.
