kvnp / sketch-to-image Goto Github PK

Sketches for Searching

A deep learning approach to make the city a better place.

We have made a project that would generate images from the provide sketches. Those sketches which are made by the mere guesses made by the eye witness and the sketch drawing artist. Thus, to give it a faster and far clear sight to those sketches, we came up with a data to generate the sharper images from those sketches. The produced images are made by using deep learning model, thus providing it a upper edge in the competing world.

Further new domains of this process have been explored by our team. The idea to identify the criminals and the missing people which get ignored and get somewhere in the dark crowdy cities. A recent case of Ropar where 33 people went missing in just 8 months. In such high crices time there is need of accurate, fast and reliable detection of people over various available media. This is quite a challenging task for the city police itself. Thus, an efficient way is to bring the machine learning in the ground.

No doubt a great work has been done in image generation. The images were being generated by computers and sold in auctions at such high rates. But in such a trend the cost-effective use for the betterment of the human society was left behind. Thus, the image generation could be used in the cities like Chandigarh. The cities that are aware with the technology and its advantages. Here it can be used for the following reasons:

1. ​ Detection of criminal people.
2. ​ Finding the lost people.
3. ​ Used in shops as to make the citizens feel safer and near to the organisations.
4. ​ Can be portable and accessed just by a click.
5. ​ Moreover, as the technology becomes common it would become more cost efficient.

We have come up with the web-based system. It is a very cost effective and scalable system. As just a single server or say a mainframe would be suffice for thousands of its users than individual using his/her own hardware and doing the same repeated work again and again. Moreover, the system is robust. As the backend of the system is designed in Django, which is the python based frame-work. It is capable of handling thousands of requests per second. Which is capable of handling high loads of requests from its users. The front GUI being designed in HTML and CSS make it light weight and scalable and accessible on various hardware easily.

According the news article of HindustanTimes the cameras in Chandigarh, that counts to 14,100, CCTV cameras are there. From which maximum of them are traffic cams used at the traffic signals. The only use of these cameras is to keep an eye on the traffic rule violation. Moreover, these cameras are only scanned after the accidents happen as to get the proves of the accidents. Such a great resource is not being utilised to its full extent. Thus, our idea could easily fit in and take the security of the city as well as the technology to a greater height.

Our model is economically and technologically feasible. We have successfully implemented U-net with ResNet as backbone for image to image translation. The sketches are converted to realistic graphic photo which further used to track down a criminal or a missing person. The model could easily be deployed in embedded systems also which makes it portable and easy to use. We used Keras framework with TensorFlow at backend therefore we can easily convert the model into graph representation and it is system independent. This model could easily be put on work on mobile devices. Which increases it’s accessibility. We will put this model in a mobile app and connect that app to our server architecture. The police man having access to the app could just click the photo of person and could get the result whether he is required by any police station. Our solution aim is to make our city safer and our police stronger. Our police should be a few steps ahead of the criminals always.

Model Architecture

We used U-net with ResNet as backbone. U-net work on the concept of encoder and decoder. It has skip connections which help it to learn more features. It has proved great in many real world problems.

Dataset

We have the Celeb dataset download from kaggle : https://www.kaggle.com/jessicali9530/celeba-dataset

Then using image processing techniques we converted those images into sketch. The code related to this is in image2sketch.ipynb file.

Flowchart

This is the rough working steps of our architecture. First we perform image preprocessing like removing the noise on the input image. Then we pass that image to our model. Our model will output a photo graphic image. After that we can present that image in any required format.

Results

Our model was trained on a Nvidia GTX 1050 Ti for the whole night which is around 7-8 hrs and performed very well. Below are some of the results it has shown.

Conclusion

Technology is so strong today that we only need ideas and we can develop anything. Deep learning is at his best pace right now. Great work is released everyday. We can use this to keep our police forces ahead of criminals. Our method we can generate the real image of the person and then we can track him easily.