Fast and High-Purity Seed Sorting Method Based on Lightweight CNN

• Published in: Bio-inspired Computing: Theories and Applications pp. 607 - 615
• Main Authors: Luan, Zhengguang, Li, Chunlei, Ding, Shumin, Guo, Qiang, Li, Bicao
• Format: Book Chapter
• Language: English
• Published: Singapore Springer Singapore
• Series: Communications in Computer and Information Science
• Subjects: Convolutional neural network; Heterogeneous convolution; Seed sorting
• ISBN: 9789811534140; 9811534144
• ISSN: 1865-0929; 1865-0937
• DOI: 10.1007/978-981-15-3415-7_51

Seed purity is an important indicator of seed quality. Seed sorting has been extensively studied and is considered as image classification setting with the goal of distinguishing between normal and abnormal seeds. Traditional, the classification of normal and abnormal seeds is achieved using the machine visual features of the seeds. In recent years, convolutional neural network has shown excellent performance in image classification tasks. In this work, we mainly focus on the computational efficiency and the classification performance of the network. Then, we developed a lightweight convolutional neural network to achieve fast and high-purity seed sorting. A lightweight and fast CNN model is constructed by using heterogeneous convolution layer instead of standard convolutional layer. Specially, we compare the standard convolution network and the heterogeneous convolutional network to measure the performance of methods on sunflower seeds dataset. The proposed sunflower seed sorting method based on heterogeneous convolutional network is robust to classify the seeds and the accuracy of data can reach 98.6%, which FLOPs are half the original standard convolution. Compared with the other state-of-art methods, this method has higher performance and lower computational complexity.