Alifya Febriana

Green coffee beans are among the most globally recognized agricultural products. However, the conventional practice of manual sorting has been riddled with inaccuracies due to human error. Recent advances in 3D monocular imaging, which employ monocular depth estimation to generate 3D point clouds, transform cameras into pseudo-LiDAR sensors. This method proved to be more accurate in our study, delivering detailed information about bean size, shape, and potential flaws, thereby presenting a potential foundation for developing a green coffee bean sorting system. In this paper, we propose a more refined pseudo-LiDAR model that is not only simpler but relies solely on a monocular camera with basic calibration parameters. Moreover, our model generates pseudo-LiDAR directly from depth maps created by a single monocular camera. Through our quantitative analysis, we have achieved highly satisfactory results in terms of distance metric similarity and Intersection over Union (IoU), validating the efficiency and precision of our model. To the best of our knowledge, this is the first study that explores the application of LiDAR-based detection algorithms for coffee beans.

Coffee beans are one of the high-value commodities in Indonesia, but the sorting method for the quality of coffee beans still uses visual methods and sieves with mechanical machines. This study aims to provide an alternative to classifying coffee beans using Deep Learning algorithms, such as the ResNet-18 and MobileNetV2. The dataset used is USK-COFFEE which consists of two classes, the normal coffee bean class, and the defect coffee bean class. The models were trained using the balanced and imbalanced datasets, meaning that two scenarios were performed. From the results of this study, the accuracy value for the ResNet-18 model was 91.50% when trained on the balanced dataset and 90.87% on the imbalanced dataset. It outperformed the MobileNetV2 architecture, which only achieved 0.8125 and 0.8112 accuracy scores when trained on the balanced and imbalanced datasets, respectively.

Coffee is one of the plantation commodities that plays a big role in the world economy. According to the classification of coffee, each type of coffee has various shapes and textures. Traditional human visual sorting of coffee beans is time-consuming, labor-intensive, and may result in low-quality coffee due to work stress and exhaustion. The contribution of this paper is twofold. First, a new dataset, called USK-Coffee, which contains a total of 8.000 images and is divided into 4 classes, is created and made publicly available. To the best of our knowledge, the USK-Coffee dataset is currently the most comprehensive green coffee bean dataset. Second, this study aims to offer a lightweight and understandable intelligent coffee bean sort accurately system that uses deep learning (DL) to assist farmers in sorting green bean arabica by variety. To be specific, this paper presents a baseline for classification performance on the dataset using the benchmark deep learning models, MobileNetV2, and ResNet-18. These models achieved an average classification accuracy of 81.31% and 81.12%, respectively. The dataset is available at: http://comvis.unsyiah.ac.id/usk-coffee/

Tomatoes are a horticulture product with the potential to be developed since they have a high economic value and are in high demand by industry and consumers. Tomato plants, on the other hand, still need to be handled carefully in order to boost yield harvest. Furthermore, susceptibility of the tomato plants to Septoria leaf spot disease, which arises due to Septoria Lycopersici Speg fungal infection, is being one of the challenges in escalating tomato production itself. Regarding the problem, this study aims to detect Septoria leaf spot on tomato plants by developing a tool utilizing deep learning and Convolutional Neural Network (ConvNets or CNN) model. CNN model conducted in this work is a supervised learning technique that extensively operated for solving linear and non-linear problems. Moreover, Raspberry Pi microcontroller and Intel Movidius Neural Computing Stick (NCS) are employed in this work, in order to accelerate the computing process and to ease the detection process considering its portability, speed, and accuracy. The precise detection range is from 84.22 percent to 100 percent, with an average accuracy rate of 95.89 percent.