Neural network-based color computer vision for grading tomatoes (Lycopersicon esculentum)

The authors develop a computerized color image analysis procedure and neural network model (NNM) to automate the classification of maturity of fresh tomatoes. Automating the classification procedure will help reduce errors of human graders who compare tomato color with a color chart. The chart is a USDA standard that human graders use to classify fresh tomatoes into six maturing stages: Green, Breakers, Turning Pink, Light Red, and Red. The repetitive procedure of visually comparing colors is prone to human errors and subjectivity, while the numbers of wrong classifications increase with time as human graders experience eye stress, boredom and tiredness. The accumulated system uses a computer color vision as its artificial 'eye' and a NNM as its artificial 'brain'. The authors setup a computed-mounted digital camera that captured 6,000 color images of locally grown and harvested tomatoes equally representing the six maturity stages. The authors classified each tomato according to the majority grade given by five-expert graders from a local commercial farm. The authors developed a color image analysis procedure to extract the red, green and blue (RGB) spectral values of the captured images. The authors designed a tomato maturity classifier based on a 3-layer NNM that uses the RGB spectral values as inputs. The NNM was trained via the feed-forward, back propagation algorithm using 4,200 images as training data and 600 image as test data, equally representing each maturity stage. To avoid model over-fitting, the authors used the NNM errors with the test data as the trainings' stopping criteria. The authors used the remaining 1,200 images to validate the model and results show that the system agreed with manual grading 97% of the time. The remaining 3% were misclassifications but within one maturity stage difference. This result shows that the authors' automatic vision system possess the same grading accuracy as the human experts but it is more efficient than the manual grading.