Image classification of forage grasses on Etuoke Banner using edge autoencoder network

Abstract

Automatically identifying the forage is the basis of intelligent fine breeding of cattle and sheep. In specific, it is a key step to study the relationship between the type and quantity of forage collected by cattle and sheep and their own growth, cashmere fineness, milk quality, meat quality and flavor, and so on. However, traditional method mainly rely on manual observation, which is time-consuming, laborious and inaccurate, and affects the normal grazing behavior of livestock. In this paper, the optimized Convolution Neural Network(CNN): edge autoencoder network(E-A-Net) algorithm is proposed to accurately identify the forage species, which provides the basis for ecological workers to carry out grassland evaluation, grassland management and precision feeding. We constructed the first forage grass dataset about Etuoke Banner. This dataset contains 3889 images in 22 categories. In the data preprocessing stage, the random cutout data enhancement is adopted to balance the original data, and the background is removed by employing threshold value-based image segmentation operation, in which the accuracy of herbage recognition in complex background is significantly improved. Moreover, in order to avoid the phenomenon of richer edge information disappearing in the process of multiple convolutions, a Sobel operator is utilized in this E-A-Net to extract the edge information of forage grasses. Information is integrated with the features extracted from the backbone network in multi-scale. Additionally, to avoid the localization of the whole information during the convolution process or alleviate the problem of the whole information disappearance, the pre-training autoencoder network is added to form a hard attention mechanism, which fuses the abstracted overall features of forage grasses with the features extracted from the backbone CNN. Compared with the basic CNN, E-A-Net alleviates the problem of edge information disappearing and overall feature disappearing with the deepening of network depth. Numerical simulations show that, compared with the benchmark VGG16, ResNet50 and EfficientNetB0, the f1 − score of the proposed method is improved by 1.6%, 2.8% and 3.7% respectively.