Now these are visual features which are based on edge detection. The higher features, the facial recognition and many more. Now when these features are extracted and then the model is created, to test this model we provide a test data which is again a set of images. In our case and using this model we get our final output in which we have the objects detected in an image as shown in Fig.
Fig. 12. TensorFlow technical architecture
Now the images are converted into a numpy array in the TensorFlow object detection so that the competition can be made easy.
We also use the TF record which is the TensorFlow record which contains the record of the image along with the tags such as you can see here, we have the person tag we have the dog tag and the horse tag. Now these tags are just labels which are provided in the input data.
Please find the solution design as shown in Fig. 13. So for that just go to github and type TensorFlow which is the official github repository of TensorFlow and inside that we have the model section. Just go to these models you can either clone the TensorFlow model or download it as per your wish.
Now the TensorFlow object detection model uses protobuf. To configure model and the training parameters. Before the framework can be used, the protobuf libraries must be compiled. Now to download protobuf, all you need to do is go to Google protobuf in github.
You can see this TensorFlow object detection API gives an accurate machine learning model description of how the objects are detected and how you have the steps for the setup.
And based on that our model will be created. Now next what we are going to do is we are going to select which model to download. So for example here we are using the faster rcnn inception v2 coco 2018 as shown in Fig. 14
So as you can see we have the test image folder here as shown in Fig. 15 and inside that you can input all the images whichever you want to test upon this model. So for example after taking the range from 1 to 8 is taking all the images named image 1 to 7. So let’s load this as shown in Fig. 16.
So as you can see we’re using the Load image into a numpy array. We are using the np.expand and finally we are using the Matplotlib to show us the results that is Labeled Processed Images.
A. Output Result (Applying deep learning on thermal imagine) are as shown below.
Fig. 17. Deep Thermal Imaging output-I where two persons are detected
Fig. 18. Deep Thermal Imaging output-II v/s visible imagery comparision where one person is detected fully hidden in vegitation(visible imagery).
Fig. 19. : Deep Thermal Imaging output-III where one person is detected through a long range thermal camera.
As you can see it identified persons as 87%, 89%, 88%, 99% sequentially as shown in Fig. 17, Fig. 18 and Fig. 19. It has provided a box; the label and the score which is the detection scored how much it is similar to all the images which have been imported in the coco data set. You can see it has not detected the car in fig. 17 as we have filtered out car object in our detection model.
B. Output Result (Superior detection in clutter)
We mentioned before that the emissions from people the black body radiation peaks at 9.5 microns. So we do a particularly good job of picking out people in a scene especially a scene which has very complex and contains a lot of clutter.
Fig. 20. Superior people detection in clutter [Visible imagery v/s Thermal imaging]
Figure 20 is basically a drive down State Street of Santa Barbara on a Friday night that happened to have a community wide bike ride. So you’ll notice the road is just completely cluttered with people. And by taking those deep learning networks that were developed originally for visible cameras doing a little training with thermal annotated images, we’ve been able to achieve pretty good results as you see here of picking out the individuals and the cars in the midst of this very cluttered urban scene.
Fig. 21. Thermal camera long range testing
And we’ve seen using this lightly trained single shot detection network. We’ve been able to detect a person up to only 20 pixels in height and if we apply up sampling to the thermal image before we apply the deep learning that work we’re able to detect a person at 16 pixels. So this corresponds to about 100 meter detection of a person with a VGA thermal camera at a 50 degree horizontal 40 degree vertical field of view as shown in Fig. 21 and Fig. 22.
Fig. 22. Thermal Semantic Segmentation
The findings(output results) of this study clearly shows that through deep learning on thermal imaging could play a better role in human search and rescue mission. This works well with adverse conditions like complete darkness, blinded with front sun rays, different types of fog, vegetation and long range.