An integrated neural network-based vision system for automated separation of clods from agricultural produce

Abstract Mechanical harvesting of onion and other agricultural produce from below ground has not been quite successful due to a large number of factors affecting the performance of harvesters. This paper discusses the integration of a neural network-based vision system with mechanical harvesters for separation of onion from soil clod to improve the efficiency of mechanical separator system. The vision system consists of a multi-layer neural network classifier that maps textural features computed from gray-scale images of onions and clods into the right object. Texture features were computed from co-occurrence matrices that specify the spatial relationship between gray-levels in the image. The textural features selected for this application consist of homogeneity, energy, contrast, and variance. The network was trained using the back-propagation algorithm. Based on this textural feature classification, the effect of changing the network configuration on separation effectiveness (SE) was also characterized. Factors including network topology and combination of textural feature measures forming the inputs of the network were systematically analyzed. It has been demonstrated that integration of the neural network vision system with mechanical harvester significantly improves the SE.