Recognize and classify vegetables in opencv and tensorflow.
The project consists a machine learning implementation, a crawler scripts to collect data, label and split the data, and an extra script to further modify the image for tensorflow machine learning format.
Tensorflow for machine learning Matplotlib for data training OpenCV for the camera module
Smart object identification: the object identification supports items in all angles, multiple items of the same type, item in a near-clear bag or container. This means the user can drop an entire bag of objects on the platform and expect the correct result to be returned. This is enabled as well as collecting and labelling most of the data ourselves, so we can choose any angle/count for the best training result.
The design (or/and reference source), training and implementation of any machine learning use in the project In order to have better and faster training results (and we don’t have gpu in training), we referenced on tensorflow and implemented a mobilenet v2 model with transfer learning. The design of the training process includes data-split, result evaluation, and final prediction. There is one file called split.py which splits four files of data [broccoli, cabbage, eggplant, carrot] into a training set and evaluation set with ratio 0.8 and 0.2. The training process is implemented in the document called mobileNetV2.py. We designed the batch size as 16, as this value is the intermediate of time cost in training, and the prediction accuracy. We split the training process into ten epochs, for each training process is within a logical time. Also, this document plots the training curve for the data in the evaluation part. In the 10th epochs, as we see the correct percentage 100% in the plotted learning curve, it is fair to say our model works in expectation.
