Document Type : Research Article
Authors
Department of Bio-System Mechanical Engineering, Bon.C., Islamic Azad University, Bonab, Iran
Abstract
Introduction
The qualitative grading of raisins, as an important agricultural commodity, plays a critical role in determining market value, pricing strategies, and competitiveness in both domestic and international markets. Conventional intelligent grading approaches have largely relied on manual feature extraction and complex image preprocessing techniques, which substantially increase computational cost and system complexity. Consequently, there is an increasing demand for automated, accurate, and non-destructive grading methods based on deep learning technologies, particularly for golden Sultana raisins.
The primary objective of this study is to develop a deep convolutional neural network (DCNN) capable of grading and evaluating golden Sultana raisins without requiring explicit image preprocessing, while simultaneously reducing processing time and computational overhead. Unlike previous studies that mainly focused on computational limitations or restricted classification schemes, the proposed model automatically extracts discriminative visual features; including texture, shape, and colour to accurately assess raisin quality and size under varying imaging conditions. This capability enhances robustness against variations in illumination, imaging angle, and background, thereby eliminating the need for complex preprocessing pipelines or manual parameter tuning.
Materials and Methods
In this study, a dedicated deep convolutional neural network, referred to as Raisin_DCNN, was specifically designed for grading golden Sultana raisin grains. The dataset consisted of 3,320 images representing four distinct quality grades. All images were acquired using a flatbed scanner under controlled lighting conditions to ensure consistency and uniform image quality.
To increase dataset diversity and reduce the risk of overfitting, data augmentation techniques were applied. These techniques included random horizontal and vertical flipping, cropping, shifting, zooming, and rotation. Notably, the proposed model was trained directly on the augmented images without applying complex preprocessing operations or manual feature extraction methods.
The Raisin_DCNN architecture was purpose-built to effectively extract visual features from each raisin. It comprises multiple convolutional layers with small kernel sizes, followed by MaxPooling layers for spatial dimensionality reduction, and fully connected layers at the final stage. ReLU activation functions were employed to facilitate the learning of nonlinear feature representations, while a Softmax function was used in the output layer to perform four-class classification. The architecture was carefully optimised to achieve an appropriate balance between classification accuracy and computational efficiency.
The dataset was divided into training, validation, and test subsets. Network training was performed using the Adam optimisation algorithm in conjunction with the categorical cross-entropy loss function. Learning rate scheduling and early stopping strategies were implemented to improve convergence and prevent overfitting. Hyperparameter tuning was conducted through iterative experimentation based on validation performance.
For comprehensive performance evaluation, the proposed Raisin_DCNN model was compared with transfer learning approaches based on pre-trained InceptionV3 and InceptionResNetV2 architectures, as well as classical machine learning methods employing histogram-oriented gradients (HOG) and scale invariant feature transform (SIFT) feature extraction combined with support vector machine (SVM), k-nearest neighbours (KNN), and logistic regression (LR) classifiers.
Results and Discussion
Model performance was assessed using accuracy, precision, recall, and F1-score metrics. Experimental results demonstrated that the Raisin_DCNN model achieved an overall classification accuracy of 96%, outperforming both the transfer learning models and classical feature-based approaches. This superior performance highlights the effectiveness of the proposed network in automatically learning discriminative features directly from images of raisins. Furthermore, the elimination of complex preprocessing stages significantly reduced computational cost, thereby improving system efficiency and practical applicability.
Conclusion
The results of this study confirm that a dedicated deep convolutional neural network can provide an accurate, robust, and reliable solution for the automatic grading of golden Sultana raisins. The proposed Raisin_DCNN model demonstrates strong potential for real-world deployment in industrial grading systems and offers an effective approach for enhancing quality control processes in agricultural product inspection.
Acknowledgements
The authors sincerely thank all individuals and research groups who contributed to this work and provided technical assistance throughout the study. The authors would like to thank Department of Bio-System Mechanical Engineering, Islamic Azad University of Bonab Branch for their support and participation in this study. The authors also gratefully acknowledge the anonymous reviewers for their insightful comments and constructive suggestions, which significantly improved the quality of this manuscript.
Keywords
