Ensemble learning-based modeling and visual measurement of compound eye vision system

Miniaturized and lightweight visual systems with a certain degree of accuracy are a direction of visual system development. The compound-eye visual system has the advantage of small-range close-up measurement. In this paper, a spherical compound-eye visual system is built and the field of view is enlarged with a fisheye lens, which has the characteristic of large field of view. Model-based calibration is complex for compound-eye visual systems, and this paper designs a basic BP neural network to calibrate the compound-eye system. For the case of input uncertainty, i.e., not all subeyes are imaged, an integrated learning method based on subeye pairs is designed. The results show that the MAE after integrated learning improves by 38% relative to the basic BP neural network, and the average relative error of spatial line segments within the obtained data distribution improves by nearly a factor of three. Since the base learner of a single subeye pair is still better than the basic BP neural network, it is robust to input uncertainties such as subeye damage.