Iceberg head lettuce is the most popular lettuce crop in the UK, with almost 3 million lettuce heads being sold every week. The crop however is very sensitive to heat, which can prevent the formation of tight heads, and is highly susceptible to internal and external breakdown, infection, and pest attacks.
Since harvesting the crop is currently done by hand, and sourcing labour is becoming more difficult, manual harvesting of iceberg lettuce is slowly becoming an increasing economical liability.
There is a need for an automated harvesting system using computer vision.
A Convolutional Neural Network (CNN) is comprised of a number of convolutional nodes performing non linear operations followed by one or more fully connected layers. Each node convolves its input tensor with trainable filters, applies element-wise ReLU non-linearity to the convolution outputs and performs a non-overlapping max-pooling + sub-sampling step.
Each of the inputs has a weight that the network is then able to learn through an algorithm known as backpropagation, by minimising a softmax loss function using stochastic gradient descent. This model was trained on 1242 images of mature iceberg head images obtained from G's lettuce farms in the UK.