Python Based Defect Classification of Theobroma Cacao Bean using Fine-Tuned Visual Geometry Group16

The study aims to classify cacao bean defects based on the captured image using vgg16. Seven classes of cacao beans were gathered including broken, cluster, flat, germinated, good, insect and moldy. One hundred images per class were captured using an enclosed capturing box with c920 Logitech camera inside and LED as light source. Image augmentation was done to increase dataset. Transfer learning technique was implemented by utilizing the pre-trained vgg16 model architecture adding 10% Dropout after FC2 layer and using default weights of several layers through fine-tuning. Three methods of fine-tuning were conducted by freezing the convolutional blocks. Performance of the trained model using several optimizers (such as Adam, RMSprop and SGD) and loss functions (such as categorical crossentropy and mean squared error) were analysed. The effect of the no. of epochs as well as different learning rates during training was considered and checked. The metrics used in choosing the model were based on the confusion matrix. The chosen model is using vgg16 architecture with 10% dropout + adam optimizer + 0.0001 learning rate + categorical crossentropy loss function run in 20 epochs. It has 95.33% average accuracy. The model was embedded in a processor for actual testing. It has an accuracy of 97.29% based on the actual testing on prototype with 37 testing samples.