Tiller number is a key trait to characterize the growth of winter wheat, which is of great significance for seedling condition monitoring and yield prediction of winter wheat. Given the complicated data acquisition and thelarge volume of the estimation model, an estimation method for the tiller number of winter wheat based on visible light images and lightweight convolutional neural networks (CNNs) was explored. It can realize nondestructive and rapid estimation of tillering numbers and can be embedded into mobile terminal devices. It can realize nondestructive and rapid estimation of tillering number of winter wheat and can be embedded into mobile terminal devices. Based on these data, lightweight CNNs MobileNetV2, SqueezeNet and ShuffleNet were used to construct the estimation model of tillering number of winter wheat. The optimization comparison test was conducted, and the comparison test was conducted with the estimation model constructed based on non-lightweight neural network AlexNet and ResNet series models. Comparison tests were conducted with estimation models constructed based on non-lightweight CNNs, AlexNet and ResNet series models. In addition, the robustness of the tillering number estimation model under different plant densities and the generalization ability of data in different growing seasons were verified. The results showed that the determination coefficient (R2) and normalized root mean square error (NRMSE) of the estimation model based on MobileNetV2 were 0.7 and 0.2, respectively, which showed the best performance among the three lightweight CNNs. The volume of the winter wheat tillering number estimation model constructed based on non-lightweight CNNs was 2.3~16.1 times that of the winter wheat tillering number estimation model constructed based on MobileNetV2. Compared with the non-lightweight CNNs, the estimation model based on MobileNetV2 had better R2 and smaller volume, which was suitable for embedding into mobile terminal devices. According to the visible image data set divided by three plant densities, 120, 270 and 420 plants ·m-2, the value of R2 of the estimation model based on MobileNetV2 were 0.8, 0.8 and 0.7, respectively, showing robust performance. For visible images of two growing seasons, the estimation model based on MobileNetV2 improved R2 by 2 times, and NRMSE decreased by 7.6% through transfer learning, showing good adaptability to seasonal differences of data and reflecting the generalization ability of the model. Therefore, based on visible light images, the lightweight CNNs estimation model can meet the tillering number estimation of winter wheat and provide an accurate, robust tool that can be embedded into mobile terminal devices for winter wheat growth observation and field agronomic measures management decisions.