This repository aims to implement a VGG19 with tensorflow . it gives a pretrain weight (vgg19.npy), you can download from here.The training file contains 668 images(223 fire images and 445 non-fire images). The testing files only have 50 fire images and 50 non-fire images.Such few images impacted the precision and recall,which have lower results.We sincerely suggest you to download more images to train. We built this VGG19 with Tensorflow and Keras in Windows , it's very convenient for most of you to train the net.
- Python 3.5
- TensorFlow 1.0
- Numpy
- fire and non-fire images here
- image_generator: We abandoned the image iterator used in previous Alexnet, and created a new gennerator with tensorflow queue.
- In this VGG19 Net, we used precision, recall and f1 to measure the performance of the net, those parameters are always used in Statistics.
- The vgg19 net and datagenerator files have been builded, you don't have to modify it. But if you have more concise or effective codes, please do share them with us.
- finetune.py is aimed to tune the weights and bias in the full connected layer, you must define some varibles,functions,and class numbers according to your own classification projects.
- precision = 60%
- recall = 95%
- f1 = 72%
We choosed some pictures from the internet to validate the VGG19 Net, the sunset images or warm lights images are difficult to identify, and some non-fire images misidentified. See the results below: