CNN-based Plant Growing Stage Classification

Ho, Zhao Jun (2019) CNN-based Plant Growing Stage Classification. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text HoZhaoJun.pdf Restricted to Registered users only Download (2MB)

Abstract

These project going to present the process of constructing an optimize architecture of the CNN models for the plant growing stage classification task using the raw greyscale plant image. In order to construct the optimize CNN models, four parameters have been manipulated, which are the number of filter in each convolutional layer, number of convolutional layer, number of neuron in each fully-connected layer and the number of fully-connected layer. Each of these parameter will be tuned independently and the parameter value with the highest mean test accuracy will be served as the optimum setting for the next parameter tuning process. Furthermore, this project also try to investigate the possibility for improving the classification success rate by adding additional information into the input image. The potential useful information are mutant type, number of leaf, total leaf area, mean leaf area and the standard deviation of leaf area of the plant, each of these additional information will be tested separately and the best mean test accuracy for each additional information will be recorded. Those additional information that are able to improve the classification success rate compare to the raw greyscale plant image will be treated as useful information. On the next part of the experiment, all the useful information will be added into the raw greyscale plant image and construct another optimize CNN models to identify the final best mean test accuracy of the CNN model. The best mean test accuracy for the raw greyscale plant image CNN model is 71.82%. All the additional information are able to slightly improve the classification success rate of the CNN model and they have the best mean test accuracy of 73.03%, 74.85%, 74.55%, 73.79% and 72.88% for the mutant type, number of leaf, total leaf area, mean leaf area and the standard deviation of leaf area respectively. After adding in all the useful information, the best mean test accuracy has been improve to 81.92%. These shows that by adding in useful information will help the CNN model to have a greater success rate in plant growing stage classification task.