Utilizing Different Edge Detection and Preprocessing Techniques to Improve the Accuracy of Durian Cultivar Detection using Convolutional Neural Networks

Using machine learning to recognize images of fruits or plants is a very practical and focused application. When using machine learning principles to classify specific fruit varieties, as opposed to traditional approaches, a higher level of accuracy is promised but also requiring more training data. Particularly in situations where different cultivars of the same fruit family are valued differently, the capacity to recognize and classify diverse cultivars of a fruit family is crucial. The objective of this study is to develop a practical method for classifying the many cultivars of Durio zibethinus, also known as Durian, based on the visual traits of the crop using Convolution Neural Networks (CNNs) to improve the accuracy and speed of cultivar recognition. Utilizing a database of images with tagged Durian cultivars. Images from two high-value classes of Durian cultivars are utilized to train the neural network. The first phases in the research project involve pre-processing and converting the images using Canny, Laplacian, Sobel on X-axis, and image binarization. Next up are model training, model validation, and network architecture development. Finally, they were combined as an ensemble machine learning method using majority voting to examine the potential for enhancing prediction performance. Laplacian edge detection was found to yield the best results of the four studied approaches, while ensemble methods gave more accurate results than a single method.