Size and weight estimation of Konjac with occlusion classification and deep learning
Kabin Yok-In, Prawit Buayai, Xiaoyang Mao · Measurement · 2025
This paper proposes a novel machine vision-based technology using a single RGB image to estimate the size and weight of konjac, a popular Asian crop, in load containers. Konjac’s irregular shapes and frequent occlusions from overlapping produce make traditional measurement methods impractical and labor-intensive. To address this, we present a multi-stage system combining a fine-tuned DetectoRS model, custom occlusion classification, and contextual dimension estimation. This approach significantly reduces the Mean Absolute Error (MAE) in dimension estimation for occluded Konjac from 2.38 to 1.1 cm and 2.66 to 1.38 cm for major and minor axes, respectively. Consequently, our method improves the accuracy of weight estimation by decreasing the Mean Percentage Error (MAPE) from 35.8% to 9.33% – a 72.3% reduction in error rate. Importantly, this approach is not limited to konjac; it can be applied to a wide range of crops experiencing occlusion in packed conditions. This versatility makes it a valuable tool in precision agriculture, offering a scalable solution for accurate weight and size estimation across various agricultural products, enhancing crop quality assessment.