The geometry and dimensions of atomic force microscopy (AFM) probe tips are critical parameters for precise measurement of surface micro-nanostructures and accurate characterization of local physicochemical properties. While conventional blind tip estimation methods based on mathematical morphology can evaluate tip geometry solely from scanning images, they typically provide upper-bound estimates rather than true tip dimensions and suffer from significant sensitivity to scanning noise, resulting in insufficient measurement accuracy. To overcome these limitations, this study proposes a robust convolutional neural network ( CNN) with an encoder-decoder architecture for stable and accurate AFM tip characterization. During supervised learning, a training dataset was generated by simulating scanning images of nanoparticle structures with varying radii and densities through mathematical morphology dilation operations, representing tips with predefined dimensions. The network parameters were optimized using mean absolute error as the loss function. Experimental results demonstrate that the CNN model achieves accurate tip radius predictions for scanning images when the tip dimensions fall within the training range. However, the model exhibits reduced accuracy for tip sizes outside the training distribution. Notably, the model′ s predictive capability is significantly enhanced through noise-augmented training data, enabling precise tip dimension estimation from noisy scanning images without requiring additional denoising procedures. Validation using actual AFM scanning images confirms the method′s effectiveness in practical applications. Furthermore, simulations and experimental data verify the method′s extensibility for processing tip-effect-distorted images.