With the rapid development of industrial automation, PLC-controlled machine vision systems have been widely applied in the classification of granular agricultural products. However, traditional PLC control systems face bottlenecks in computational efficiency and classification accuracy when processing complex image data. To address this issue, this study proposes an optimization method combining the Particle Swarm Optimization (PSO) algorithm with Convolutional Neural Networks (CNNs) to enhance the classification performance and stability of the system. The PSO algorithm is employed to optimize the structure and hyperparameters of the CNN, ensuring maximized classification accuracy and improving the system’s real-time performance and robustness. Experimental results demonstrate that the PSO-optimized CNN model (PSO-CNN) exhibits excellent performance in the classification of granular agricultural products, achieving an accuracy rate of 94.1%, significantly higher than the 85.2% of traditional CNNs. Additionally, the error rate of PSO-CNN is 5.9%, much lower than the 14.8% of CNNs, showcasing its advantages in complex data processing. Furthermore, PSO-CNN maintains a high level of classification accuracy under noisy conditions, particularly with a Gaussian noise level of 0.1, achieving an accuracy of 89.3%, which proves the robustness and stability of the model. Overall, the PSO-optimized CNN model demonstrates high accuracy and good real-time performance in the classification of granular agricultural products, effectively addressing complex conditions in practical applications and possessing strong practical value and promotion potential.