Data Augmentation for Deep Learning based Cattle Segmentation in Precision Livestock Farming

摘要

Accurate segmentation of cattle is a prerequisite for feature extraction and estimation. Convolutional neural networks (CNN) based approaches that train models on the largescale labeled datasets have achieved high levels of segmentation performance. However, pixel-wise manual labeling of a cattle image is challenging and time consuming due to the irregularity of the cattle contour. In this regard, data augmentation for deep learning based cattle segmentation is required. Our proposed data augmentation approach uses random image cropping and patching to expand the number of training images and their corresponding labels, then, a state-of-the-art deep neural net is trained to segment cattle images. Here we apply these techniques to images of cattle in a feedlot environment. Our data augmentation-based approach segmented cattle from a complex background with 99.5% mean Accuracy (mAcc) and 97.3% mean Intersection of Unions (mIoU), improving current techniques including a combination of random flipping, rotation and color jitter.