| Summary: | There are technologies available in research in order to have inspection on macro and micro cracks on eggshell. However, there are still some difficulties when coming to inspection on translucent areas where before the micro-cracks happening. Transfer leaning using pre-trianed neural network is used at minimized computational resources while having a very high efficiency in classifying the eggs into three classes which are good, bad and unknown. Alexnet, Resnet and Inception of different architectures are compared to compute respective accuracy. It proved that the Alexnet gives highest predictive accuracy which is 96.80%, followed by Resnet, 93.15% and Inception, 90.16%. Results obtained from Alexnet is used to do statistical analysis such as ANOVA and student-t test to measure statistically significant differences between the means of accuracy from training set and testing set of image data. Visualization on channel along with activation strengths allow to know how a network learn to classify an egg with the help of Pareto chart. The deep dream images are generated by referring to the generation of images that produce desired activations.
|
|---|