To address the problems of a large number of parameters, low real-time performance, and poor anti-interference ability of traditional semantic segmentation networks, a visual detection method of suspension gap based on column-oriented semantic segmentation is proposed. The method defines gap detection as finding the set of disaggregated locations of gaps in the middle of the image and simplifies the classification problem to reduce the computational complexity. Firstly, the structure of the visual gap detection-based suspension system is designed. A gap detection semantic segmentation network (GMSSNet) is designed based on position selection and classification in the column direction. The number of model parameters is further reduced by using a 1×1 convolution plus column-wise sub-pixel convolution layer module instead of a fully connected layer. Then, the suspension gap sample set is constructed and the training environment is configured. The designed GMSSNet model is tested for anti-jamming ability, ablation experiments, and closed-loop suspension experiments, respectively. The experimental results show that the GMSSNet model has high detection accuracy, the maximum detection error is ±0.1 mm and the linearity is 0.5% F.S for normal levitation gap detection samples. The maximum detection error of the network is ±0.15 mm and the linearity is 0.75% F.S in the presence of offset or specific occlusion. The closed-loop levitation experiments show that the levitation gap detection accuracy and speed of the GMSSNet model meet the requirements of the suspension system.