As an efficient geophysical exploration method, controlled source electromagnetic exploration is often limited by factors such as electromagnetic signal attenuation and scattering, as well as formation complexity, and the resolution of controlled source electromagnetic 3D inversion of deep targets is often limited. In order to improve the resolution of deep exploration, this study proposes a deep learning method that fuses CBAM (Convolutional Block Attention Module) and CNN (Convolutional Neural Network). CBAM can effectively learn the attention between feature channels and spatial positions, thereby enhancing the model's ability to extract important information. The CBAM module is combined with CNN to further extract important features of deep structures through multi-layer convolution and pooling operations, thereby improving the resolution of inversion. Firstly, based on the contraction integral equation developed by the CEMI of the University of Utah combined with the regularized iterative optimization method of focusing on the stable functional, the three-dimensional electromagnetic forward modeling and inversion calculation of the deep target under inhomogeneous background are carried out. Then, the inverted three-dimensional conductivity distribution data is sent to the CBAM-CNN network for training and optimization, which can improve the resolution of deep exploration while maintaining efficient calculation. Compared with the traditional methods, the CBAM-CNN method proposed in this study can significantly improve the resolution of deep controlled source electromagnetic exploration, restore the underground structure more accurately, and obtain more refined underground resource information. This method can provide a new deep learning method for the accurate positioning of deep geological structures and resources, and has important application value.