Multilevel thresholding is a widely used method in image segmentation. However, the traditional methods are costly to obtain the optimal thresholds through exhaustive search. The Nature-inspired algorithm is a gradient-free optimizer that overcomes these shortcomings and generates the best thresholds with high quality and efficiency. For this purpose, this paper suggests an improved arithmetic optimization algorithm with federated opposite learning for multilevel thresholding image segmentation, namely FOL-AOA. In this method, the federated opposite learning strategy is incorporated to avoid the particles being trapped into local optimal and increase the population diversity. The cross-entropy is employed as the objective function minimized by FOL-AOA. To assess the performance of the proposed method, we considered the use of a variety of benchmark images under different threshold levels and compared them against five predecessor approaches. The obtained results manifest that FOL-AOA outperforms the comparison methods in terms of the fitness values as well as two image quality indicators such as PSNR and SSIM.