Oil palm fresh fruit bunches ripeness classification with color and texture features extraction

Abstract

High quality palm oil is critical in ensuring Malaysia’s competitiveness in the industry. Studies have shown that there is a significant relationship between the quality of palm oil produced and the ripeness of the fruits used in producing the oil. Correct ripeness of the fresh fruit bunches (FFB) produces higher quality and more oil content. Unripe FFB produces less oil and overripe FFB produces oil of low quality. According to Malaysian Palm Oil Board (MPOB), the main factors that determine the ripeness of the oil palm FFB are its colour and the number of its loose fruits. The purpose of this study is to improve the FFB grading accuracy at the plantation field using an image processing approach, which currently the grading is done manually by human graders. Most research in this area are based on FFB images taken at controlled environments with perfect lighting and taken using high-quality cameras. This is quite impractical as the FFB quality must be determined at the harvesting sites before they are sent to the mill for processing. In this research, images used were taken using a common handphone camera at the plantation site after the FFBs were harvested. The analysis combines both the colour and texture features of the FFB, which the colour features of 400 FFB images were analysed using different colour channels, as each colour channel has different responses to the sunlight. The Red Green Blue (RGB) colour channel of the images went through image processing with colour conversion into YCbCr (Y(Green), Cb (Blue), Cr (Red)), which this colour selection is empirically recommended for outdoor images. The results showed that by converting the images to YcbCr, the accuracy of the grading is more promising with increasing of 40% compared to using RGB colour channel. Deep learning was used to identify the colours and textures. In view of small dataset of this study, additional image pre-processing techniques with colour conversion and resizing on raw image data were processed before using them as inputs to the Convolution Neural Network (CNN). The findings showed that with the pre-processing step, even with small dataset, the CNN can produce good FFB ripeness classification with accuracy 74%, with acceptable inferencing time of 3 seconds. In sum, the findings contribute to the development of technology to identify the FFB ripeness in the real plant harvesting scenario by camera.