Adaptive high-distance RGB imaging for accurate dairy cow feed intake estimation

This study proposes a method for estimating the total feeding amount in the feeding area of dairy cows based on Red-Green-Blue (RGB) images, with the aim of providing ranch management with a cost-effective and efficient intelligent measurement solution. The method utilizes a stereo camera mounted at a height of 2.95 m to capture RGB images of different feed piles, constructing a dedicated differential image dataset. In order to effectively exclude the interference of background factors in the feeding scene, we used the U2-Net network to segment these images. Furthermore, we innovatively integrate the self-attention mechanism and multiscale fusion techniques with ResNet, designing and implementing a deep learning model for estimating the total feeding amount within the camera's field of view. The experimental results show that, within the 0 to 10 kg range, the proposed method achieves the mean absolute error of 0.3487 kg and the root-mean-squared error of 0.4456 kg, outperforming commonly used methods in real-world scenarios.