Cashew nut size classification using hybrid learning techniques

Abstract

The research problem originated from the limitations of traditional cashew nut size classification methods, which result in high labor costs and inconsistent product quality, affecting the competitiveness of Thailand’s cashew nut industry in the global market. The research objective was to develop a high-precision cashew nut size classification system by integrating deep learning and machine learning techniques. The proposed methodology involves a hybrid cashew nut size classification system. First, a deep learning technique utilizing YOLOv5 was applied to detect cashew nuts from images, with experiments conducted to determine the optimal and most efficient parameters. Subsequently, machine learning using a Support Vector Machine (SVM) was employed for size classification based on dominant morphological features, including width, length, and area. The SVM model was evaluated using four kernel types: linear, radial basis function (RBF), polynomial, and sigmoid. The results indicated that the detection model performed optimally when using 400 epochs and a batch size of 64, achieving a precision of 0.9817 and a recall of 0.9706. Furthermore, the size classification model achieved perfect performance, yielding precision and recall values of 1.0 across all tested kernels. However, the overall system performance still had minor limitations due to occasional errors in the detection step. In conclusion, this system can improve the accuracy of cashew nut size classification and enhance the competitiveness of Thailand’s cashew nut industry in the global market by reducing labor costs and improving product quality consistency. Future developments should focus on refining the detection process to further enhance the overall system efficiency.