Transfer learning using mobileNet for rice seed image classification

Abstract

Rice is the world's primary source of carbohydrates, especially in Asia. Quality rice requires good seed breeding. In this research, we classified rice seeds. The experiment using public data consists of five classes. Each class contains 2,000 images. The total amount of image data is 10,000. Classification uses mobileNet, which consists of 13 depthwise separable convolutions consisting of depthwise (DW) and pointwise (PW) convolutional layers. Each DW and PW is followed by batch normalization and Rectified Linear Unit activation. At the end, there is Global Average pooling and two dense layers. The trial uses transfer learning with initial weights from imageNet. The first to twelfth convolutional layers freeze. That is, they do not train the weights in them. On the 13th or last convolutional layer, fine-tuning is carried out. Experimental data is divided into training, validation, and testing. The testing results show that accuracy is 99.55%, precision 99.55%, recall 99.08%, and f1-score 99.31%.