For catalytic materials, the size and shape of nanoparticles and other structural information have an important impact on the catalytic performance. The identification and statistics of nanoparticles based on transmission electron microscope images are the main means to obtain these information. In this paper, a deep separable convolutional U-Net architecture based on deep learning is proposed. Taking the core-shell nanomaterial as dataset, cross-entropy loss function, weighted cross-entropy loss function, IoU loss function and Dice loss function are adopted as optimization objectives to train the network respectively. The segmentation results show that the performance of IoU（Intersection over Union）loss function and Dice loss function is better for the dataset of core-shell structure nanoparticles with unbalanced positive and negative samples. Finally, the trained network is used to segment and conduct statistics on TEM images to obtain structural information such as particle size and perimeter distribution, which provides feasibility for the application of deep learning in the field of catalytic materials.