The safe exploration of Small Celestial Bodies (SCBs) requires continuous human supervision and participation, which limits the number of missions that can be realized. If spacecraft were endowed with greater onboard autonomy, exploration could be carried out more efficiently, resulting in broader coverage and increased scientific output. By reducing the need for extensive human involvement, these missions could be conducted more frequently and with greater ease. This paper presents an AI-based, especially Convolutional Neural Network (CNN)-based, pre-processing for improving the navigation concept of the previously finished Astrone project, which introduces an exploration concept that suggests using autonomous hovering relocation maneuvers for exploration spacecraft on Small Solar System Bodies (SSSBs), such as comets and asteroids to enhance their mobility. Our improvements support autonomous navigation concepts by extracting relevant features from the spacecraft's irregular and unstructured environment, such as the comet 67P/Churyumov-Gerasimenko's surface, captured by a 65 • ×65 • wide-angle 2D camera and a corresponding Light imaging, Detection And Ranging (LiDAR) sensor. This paper focuses on predicting Volume Center Points (VCPs) of different-sized stones in the spacecraft's camera frame as landmarks by CNNs. These VCPs are treated as key-points for local navigation, such as point cloud registration and place recognition. A simplified dataset for our specific scenario was generated to train and test different CNNs, such as the U-Net, DEFU-Net, U 2-Net, and the Deeplabv3+ with certain ResNet and ResNet-RS backbone net architectures. The CNNs were trained and tested for semantic segmentation tasks and VCP detection for different-sized stones. Further, the Deeplabv3+ with a ResNet-RS50 backbone net was extended by three separate Multi Perceptron Layers (MLPs) to predict the VCP coordinates independently. A Mean Absolute Error (MAE) within 0.837·σ radius could achieved, where σ radius is the stones' radius' standard deviation.