Due to the small size and high background interference of microaneurysms in retinal images, traditional methods have low detection accuracy. Currently, deep learning models mostly focus on detecting large-sized targets, which have complex structures and poor detection performance for small targets. To address these issues, a micro aneurysm detection method based on parallel multi-scale convolutional neural networks is proposed. Firstly, a corresponding relationship between the size of microaneurysms and the theoretical receptive field used for detection is established. Then, a parallel convolutional network consisted of two receptive field scales is constructed, which is based on the type and size range of microaneurysms. Finally, a training set construction and data augmentation method based on active learning is proposed to improve the detection performance of the model. The method is compared on two public datasets and one self-collected fundus dataset. The experimental results show that the method can effectively detect microaneurysms, and has better detection performance compared to similar methods for small and vascular adherent microaneurysms.