Using SURF to Improve ResNet-50 Model for Poultry Disease Recognition Algorithm

ResNet-50 is an architecture of residual network and known to have numerous advantages. However, the application of the model to the poultry domain for identifying chickens' diseases has demonstrated insufficient and overfitting results. This is due to the limitation in the training data set which comprises the whole images of chicken body, while the diseases in chickens have been known to be involved specific chicken body parts. As such, in this research work, it has been hypothesised that by pre-processing the data, specific features could be effectively identified during training. Therefore, this research uses the combination of SURF feature analysis with K-means model and then re-selects the main characteristics such as head, wings, legs, and other specific parts of chickens where the known diseases could be identified. The obtained data set was later provided into the ResNet-50 model and resulted in 93.56% accuracy, which is 20% higher than the previous research.

I. Introduction

Convolutional neural networks (CNNs) is an immutable neural network based on typical weighted architecture. It is widely used in deep belief network (DBN) [1] with a layer of deep neural networks, and many implicit layers are interconnected. The structure of the CNN model is a set of overlapping transition layers [2] and using the back-propagation algorithm. In particular, the transformation layers are used as a sliding window with parameters to get information about the features of the matrix form, commonly used in today’s image classification problems. And back-propagation is the algorithm used to calculate the gradient of the corresponding estimation function for each (weight) network parameter when going from input layer to output layer, and gradient descent used to update those parameters (1). $$\begin{equation*}y=F(x,\{W_{i}\}) \tag{1}\end{equation*}$$