SapotaNet: Paving the Way for Efficient Deep Learning with Lightweight Network Architecture

Abstract

Over the years, agricultural production in India has consistently increased. There remains a substantial gap between per capita demand and supply due to losses, including those incurred during post-harvest processing. Research in cutting-edge technologies, such as computer vision and artificial intelligence, has significantly improved fruit quality evaluation in recent years. Fruits quality evaluation by humans is an expensive and time-consuming task. Sapota fruit is a source of income for people living in tropical regions. However, having a short post-harvest life and a lack of human resources result in high post-harvest losses in the case of Sapota fruit production. Fruit grading using technology can resolve related issues. This work presents an innovative Convolutional Neural Network (CNN) model tailored for the precise grading of Sapota fruit. The proposed CNN model SapotaNet accurately grades Sapota fruit, significantly reducing the need for manual inspection and lowering post-harvest losses. This technology-driven approach offers improved efficiency, consistent grading, and higher-qual-ity output in commercial Sapota production. Since no prevailing database exists for Sapota fruit, a new dedicated database has been created, and a comparative assessment has been un-dertaken, evaluating knowledge transfer models, as well as our proposed lightweight Sapota-Net model. A total of 3,400 Sapota fruit images are assessed for Sapota fruit grading. In our experimental evaluations, our proposed model, SapotaNet, exhibited higher testing accuracy and the least inference time per Sapota fruit among all models assessed.