Automatic seeds segmentation and classification using a novel method based on pixel intensity thresholds and convolutional neural networks

Due to the computational power and memory of modern computers, computer vision techniques and neural networks can be used to develop a visual inspection system of agricultural products to satisfy product quality requirements. In this paper, artificial vision techniques are employed to classify seeds in RGB images. As a first step, an algorithm based on pixel intensity threshold is developed to detect and classify a set of different seed types, such as rice, beans, and lentils. Afterward, the information inferred by this algorithm is exploited to develop a neural network model, which successfully achieves learning classification and detection tasks through a semantic-segmentation scheme. The applicability and good performance of the proposed algorithms (effectiveness of 98.26% and a cross-entropy error of 0.0423) are illustrated by testing with real images. The experimental results verify that both algorithms can directly detect and classify the proposed set of seeds in input RGB images.