Classification of Corn Seed Quality using Residual Network with Transfer Learning Weight

Abstract

Corn is one of the main ingredients in farm animal feed. Currently, corn is preferable because widely available and cheaper in the market than others. However, it needs quality control on corn production. The company that manufactures animal feed has certain quality standards to receive corn material. On the other hand, the quality of corn produced varies greatly. Thus, quality control when receiving corn from suppliers greatly affects the quality of animal feed. The quality of feed ingredients is classified into physical properties and analytical values. Physical properties are determined so that the resulting corn can be accepted or rejected, while the analytical value is used as the basis for formulating the diet. The physical properties of corn are determined by the human senses, such as sight and smell, while the analytical value is by chemical analysis. Physical quality control by relying on human senses is certainly limited and takes time. Based on these problems, it needs to make a classification system of corn seeds automatically. This study uses corn seed images as classification data. The system uses public data from Naagar which consists of four classes: pure, discolored, silk cut, and broken. Image classification uses a Convolutional Neural network (CNN) with ResNet152v2 architecture. The hyperparameters used consist of a learning rate of 0.001, a batch size of 512, and an epoch of 25. Adaptive Moment Estimation (Adam) for the optimizer. Percentage of data training vs validation 80:20. The validation results show an accuracy of 65%, precision of 66%, and recall of 64%.