Abstract: Aiming at the problem of inefficient feature extraction in the network and the difficulty of focusing important features, an improved object detection network is proposed based on grouping residuals and transforming feature dimensions. First, a backbone network structure G-ResNet50 is introduced as a feature extraction network; then a multiple receptive field spatial attention mechanism is proposed to strengthen the attention and focus ability of important features; finally, the regression effects of different loss functions are compared and the best loss function is adopted. Experimental results show that the proposed network can balance the relationship between calculation and detection accuracy, and capture more global information at the same time, which greatly improves the detection accuracy of the network. Compared with the comparison network, the network achieves better performance on the PASCAL VOC and COCO datasets.