Performance Analysis of Deep Learning Models for Sweet Potato Image Recognition

Abstract

Current technological developments make human work increasingly automated. Computer vision has widely used deep learning to help humans recognize objects. TensorFlow is a form of CNN model that is widely used to implement computer vision. In this research, the performance of four TensorFlow models was tested to recognize yellow sweet potatoes and Cilembu, which have many similarities and are not easily distinguished by ordinary people. These two types of sweet potatoes need to be determined because they significantly differ in economic value. The four TensorFlow models tested were MobileNetV1 FPN SSD, MobileNetV2 SSD, MobileNetV2 FPNLITE SSD, and EfficientDet-D0. Based on the test results, the MobileNetV1 FPN SSD model has the best precision in all classes and has good accuracy in the yellow sweet potato class. But the performance is too lame on Cilembu sweet potato and requires the longest training time. Meanwhile, the most stable performance based on precision, accuracy, and recall is the EfficientDet-D0 model. The training process is also faster than the MobileNetV1 FPN SSD.