This study presents a Convolutional Neural Network (CNN)-based image processing system for determining the grain size distribution of coarse-grained soils, offering a faster and more efficient alternative to conventional laboratory methods. A total of 185 soil samples were collected, and high-resolution images were captured using a mobile camera. These images were analyzed using a CNN model trained to predict grain size distributions. The model’s performance was evaluated using standard error metrics: Mean Absolute Error (MAE) of 0.0645, Mean Squared Error (MSE) of 0.0645, and Root Mean Squared Error (RMSE) of 0.254. An additional set of 42 samples was collected for validation, yielding an MAE of 1.91, an MSE of 9.79, and an RMSE of 3.13. Chi-square tests confirmed strong agreement between the CNN predictions and results from traditional methods. The findings highlight the possibilities of deep learning to enhance the accuracy, efficiency, and automation of soil classification, thereby reducing the dependence on time-consuming and labor-intensive manual procedures in geotechnical engineering.