Grasp Detection is the prominent research focus in the field of embodied vision applications. In the past, grasp detection was mainly based on convolutional networks to extract visual features. Convolution mainly focuses on the extraction of local features, while the ability to capture long-distance relationships of images is weak. Swin Transformer has gained widespread application and attention in visual tasks due to its unique hierarchy and shift window mechanism. Using the powerful global information extraction ability of Swin Transformer, a grasp detection model using encoder-decoder architecture is introduced. In this model, the encoder leverages the Swin Transformer network to capture global contextual information. A local feature extraction module was developed, which included convolutional block, CABM block, and SE block. Local features are extracted by gradually increasing the reduction in the four modules. The decoder combines the output of the feature extraction module with the output of each layer of the Swin Transformer, which further enhances the feature representation capabilities. The proposed method has been experimentally validated, achieving accuracy rate of 100% on Cornell dataset and an impressive 91.4% accuracy on Jacquard dataset. Moreover, the model boasts a remarkably compact size of just 7.33 MB.