Computer Vision News - May 2019

25 Focus on Computer Vision News In this article, we chose to use the top three classes. Choosing the parameter 'top' to be 3, means that we take the multiclass prediction of the network, which is a number associated with each class, use the three argmax of this vector. Those will be the three classes of our prediction where each class gets a score for the input image being in this class. We also implemented a simple code that by using pillow library, shows the image, write the prediction on the image and save it to the disk. At this point, we have the model in our hands, and we can use it to classify images. Let’s see some results: Dog breed test usually requires a blood test, which might cost a significant amount of money. In this example we will show how to make it much cheaper. ImageNet dataset contains about 1000 classes some of the classes are objects, some are plants, some are animals, and some are animal species and subspecies. In this example, we will classify dogs into breeds by using the 120 classes of different breeds of dog that are available in ImageNet. The output of the network is a vector with 1000 entries, each of the entries corresponding to a specific class. Our prediction for the dog's breed will be the 3 highest ranking classes. Below are some results of our model. The network top 3 predictions and the score for each class are listed at the top left of each image. One can appreciate the accuracy by examining our visual results. Obviously, for more accurate classification we will need to develop a much more complex model; however this is a good start! This is a good start! Keras Pre-trained Models