Lunar crater detection is of critical importance for planetary geology and mission planning. In this work, a Faster R-CNN model based on a ResNet-50-FPN backbone is used that incorporates a Region Proposal Network (RPN) with multi-scale anchor boxes (8–256 pixels) and ROI heads to perform accurate bounding box regression and classification. The model, which is trained on the Chandrayaan-2 OHRC training dataset of 2,266 images and tested on a testing dataset of 646 images with 7,267 labeled craters, achieves a mAP (mean Average Precision @0.5) of 0.7920 at 15 epochs, the goal being mAP@0.5 ¿ 0.5. The dataset is challenging in that craters may vary from 10–150 pixels, have sparse occurrences (1–37 craters for an image), contain annotation noise, and have varied lunar terrain textures. The pixel resolution of the dataset, 0.25 m, enables accurate diameter estimation while having a 5–10% error for big craters, supporting mission-critical mapping. The training process is divided between SGD optimization (learning rate 0.005, momentum 0.9, weight decay of 0.0005) and augmentations like horizontal/vertical flips and scale jitter to stop the model from fitting too closely to the data. Joint RPN and ROI loss aid with great convergence, discovering an average of 22.6 craters in each test image. Overfitting tiny craters indicates the need for advanced regularization. This work enhances automated lunar surface analysis, supporting lunar exploration.