A CONVOLUTIONAL NEURAL NETWORK-BASED APPROACH FOR VISUAL QUALITY INSPECTION OF READY-TO-EAT CRISP LETTUCE LEAVES

Resumo

Among the leafy vegetables, lettuce is considered a product of great importance, especially
in the context of healthy eating. The species Lactuca sativa var. crispa, or crisp lettuce, is one of the
most produced and consumed vegetables in Brazil. Packaged lettuces have represented a new marketing
model and have been gaining space in the final consumer's table. As they are minimally processed to be
ready for consumption, there may still exist points of dirt or rot in the leaves due to failures in the visual
inspection process that is conducted manually, and consequently the possibility of contamination of the
product. This work presents an approach for automatic visual inspection of the quality of ready-to-eat
crisp leaf lettuce. To this end, we employed computer vision methods and a convolutional neural
network (CNN) that was trained with two databases, one composed by 12400 sub-images (windows) of
30×30 pixels and another composed by 2560 sub-images of 50×50 pixels. These sub-images were
extracted from healthy parts of lettuce leaves and parts containing the major defects (burnt edges,
putridity, and dirt and pest infestation). In the experiments conducted, the databases were enlarged
during the training phase of the CNN employing the data augmentation technique, which increases the
training set in about 1000 times. To evaluate the proposed approach two other databases containing 4211
and 1551 of sub-images were employed. For the 30×30 and 50×50 sub-images, the average hit rates
were 96.1% and 91.6%. These results demonstrate the feasibility of the proposed approach and indicate
that smaller windows provide better CNN performance.