Convolutional Neural Network implementation for image-based Salak sortation

Salak is one of potential export commodities from Indonesia. However, the main obstacle to perform Salak export is transportation which requires rigorous selection of the fruit. In the current packaging process, this sortation is done manually. In this study, convolution neural network (CNN) is applied to automatically distinguish quality of Salak. Input used in this study is the image of salak. The process involved in this study is data collection, preprocessing, classification and testing. Preprocessing is done by cutting a region of interest (ROI) containing only salak image. Classification is done by CNN, for which to get the best accuracy of the model, existing parameters should be tested and evaluated. Testing is done for two types of model, 2-class models and 4-class models. The experiments result showed that the best accuracy obtained for 2-class model is 81.45% by using learning rate of 0.0001, a single layer convolution with fifteen filters and 100 neurons in the hidden layer. The filters' size is 3×3×3. While 4-class model obtained best accuracy of 70.71% with two convolutional layers. The numbers of filter in each layer are 6 filters with the size of 3×5×5 in the first layer and 18 filters with the size of 6×3×3 in the second layer.