Development of convolutional neural network models for evaluation of body condition scores of Holstein Friesian crossbred cows

Abstract

Body condition scoring (BCS) is an efficient tool to monitor the nutritional status of cows by subjective assessment of the amount of fat or stored energy in them. This method of scoring is cumbersome, laborious and inefficient because it takes more time to score due to increasing number of animals in modern dairy farms, besides involving high cost. Therefore, this study proposed a system based on convolutional neural network (CNN) models to automate BCS of cows by image analysis. The digitally-captured images were processed using GIMP software to subtract the background from the captured images of cows. The background-subtracted images were used to detect the edges and contours using fuzzy logic edge detection method in Matlab software. Finally, the images with body contours and edges were employed as input dataset for the development of CNN models. The image dataset was classified into two groups based on the incremental BCS system of 0.25 (CNN model 1) and 0.5 (CNN model 2). The classification accuracy of the first model for 0.25 and 0.50 error ranges was 61.41% and 80.31%, respectively. Similarly, the second model achieved classification accuracy of 81.45% and 93.54%, for the same error ranges. The CNN model 2 was relatively better as compared to the first model owing to the wider BCS range of classification.