Image processing-based transfer learning approach for classification of vegetable seeds of Bangladesh.

Abstract

Research on the agriculture sector has increased within the past few years. This study focuses on developing a transfer learning-based approach to improve vegetable seed recognition and verification, addressing the crucial demand for accurate seed classification. The agricultural economy of Bangladesh heavily relies on the quality of seeds, yet manual classification methods are time-consuming and error-prone. There are several features to classify a seed, such as color, length, texture, size, shape etc. This research evaluates the classification performance of three transfer learning models MobileNetV2, Inception V3, and NASNetMobile on a dataset comprising 6,000 preprocessed images, with each of the 12 vegetable seed varieties represented by 500 images. The research focuses on enhancing model performance through adjustments in image resolution and epoch parameters. Out of all the experiments, MobileNetV2 provides an accuracy rate of 99.50% and a loss of 1.57%, this performance is validated by the test dataset. These implications enhance agricultural seed classification by improving efficiency and accuracy.