Saffron, one of the world’s most valuable spices, is often graded based on the physical characteristics of its threads. The degree of curvature in saffron threads is a key indicator of quality. Currently, the initial visual assessment of saffron in laboratories and export companies is performed by human experts, which makes human error highly probable. This study presents an AI-based approach for classifying saffron threads into three categories: Grade 1 (no curl, low curvature), Grade 2 (curled, moderate curvature), and Grade 3 (highly curled, high curvature), utilizing deep learning techniques. Additionally, a dedicated imaging system was designed and constructed to capture high-quality images of saffron threads. A dataset of 450 saffron thread samples (150 samples per grade) was collected using this system, with a hardware platform optimized for minimal noise during image acquisition. Two deep learning architectures, Convolutional Neural Network (CNN) and Google_Net, were employed for classification. Furthermore, three optimization algorithms—ADAM, SGDM, and RMSprop—were evaluated for both networks. In the best scenario, the CNN achieved an accuracy of 72.53%, while the Google_Net, utilizing the ADAM optimizer, achieved an accuracy of 86.81%, a sensitivity of 86.7%, a specificity of 93.4%, and an F1-score of 86.8% in detecting the curvature of saffron stigmas. The results demonstrate that the Google_Net model performs effectively in classifying saffron stigma curvature. This study highlights the potential of deep learning and image processing as non-destructive methods for automating the quality assessment of saffron threads, offering a scalable and efficient solution for the automated classification of saffron stigmas.