Assessment of soybean seed quality is crucial when evaluating its potential for both nutritional consumption and propagation in subsequent agricultural cycles. This investigation presents an efficient, custom-designed Convolutional Neural Network (CNN) model using deep learning for assessing soybean seedling quality through non-destructive means. The proposed system is intended to function as an enhanced alternative to current human visual inspection methodologies. The performance evaluation and comparative analysis of the CNN model employed five datasets, comprising healthy soybeans and four distinct categories of diseased soybeans. The experimental results from the dataset demonstrated a classification accuracy of 98% achieved by the proposed model, as evidenced by comparisons with previous studies. The model exhibited a balanced performance between precision and recall across all categories (healthy, broken, spotted, skin-damaged, and immature), with an average F1 score of 0.95%. However, datasets D4 (Healthy and immature soybean) and D5 (healthy and mixed unhealthy soybean) exhibited a notable decline in test accuracy to 85%, accompanied by an F1-score of 82%, indicating a requirement for additional training data. Notwithstanding this variation, the model exceptional performance on most datasets suggests its potential applicability for real-time soybean quality assessment applications.