Claims:We claim
1. An urban surveillance system in smart cities by automated detection comprising:
which can localize license plates and recognize the letters simultaneously in a single forward pass.
2. The urban surveillance system in smart cities by automated detection as claimed in claim 1 in which the whole network can be trained end-to-end.
3. The urban surveillance system in smart cities by automated detection as claimed in claim 1 not only avoids intermediate error accumulation but also accelerates the processing speed.
4. The urban surveillance system in smart cities by automated detection as claimed in claim 1 uses Convolution Neural Network.
, Description:INTRODUCTION
AUTOMATIC car license plate detection and recognition plays an important role in intelligent transportation systems. It has a variety of potential applications ranging from security to traffic control, and attracts considerable research attentions during recent years. However, most of the existing algorithms only work well either under controlled conditions or with sophisticated image capture systems. It is still a challenging task to read license plates accurately in an uncontrolled environment. The difficulty lies in the highly complicated backgrounds, like the general text in shop boards, windows, guardrail or bricks, and random photographing conditions, such as illumination, distortion, occlusion or blurring. Previous work on license plate detection and recognition usually considers plate detection and recognition as two separate tasks, and solves them respectively by different methods.
However, the tasks of plate detection and recognition are highly correlated. Accurate bounding boxes obtained via detection methods can improve the recognition accuracy, while the recognition result can be used to eliminate false positives vice versa. Thus in this invention, proposed a unified framework to jointly tackle these two tasks at the same level. A deep neural network is designed, which takes an image as input and outputs the locations of license plates as well as plate labels simultaneously, with both high efficiency and accuracy. We prove that the low level features can be used for both detection and recognition. The whole network can be trained end to- end, without using any heuristic rule.

DESCRIPTION OF THE PROPOSED WORK

To our knowledge, this is the first work that integrates both license plate detection and recognition into a single end-to-end trainable network and solves them at the same time. The main contributions of this work are as follows:
• A single unified deep neural network is proposed, which can detect license plates from an image and recognize the labels all at once. The whole framework involves no heuristic processes, such as the use of plate colors or character space, and avoids intermediate procedures like character grouping or separation. It can be trained end- to-end, with only the image, plate positions and labels needed for training. The resulting system achieves high accuracy on both plate detection and letter recognition.
• Secondly, the convolution features are shared by both detection and recognition, which leads to fewer parameters compared to using separated models. Moreover, with the joint optimization of both detection and recognition losses, the extracted features would have richer information. Experiments show that both detection and recognition performance can be boosted via using the jointly trained model.
• By integrating plate recognition directly into the detection pipeline, instead of addressing them by separate models, the resulting system is more efficient. With our framework, we do not need to crop the detected license plates from the input image and then recognize them by a separate network.
• This method is quite different from them. The most remarkable difference is that our network can be trained end-to- end, while other methods combine results from separately trained models to obtain the final detection and recognition results. With this innovation, some pre- processing, like character detection or character grouping, are eliminated, and the intermediate errors can be avoided. The learned features can be more discriminative and lead to a better performance.

Image Capturing
we create a basic infrastructure to establish the proposed Feature of image handling in python. Prepare the python environment to access video input device, in this case a camera.

Image Frame cleaning and preprocessing

Once the image is read in python we use different image processing technique on the image. The number plate detection technique is deployed. First the image is converted to grey scale image. This means the pixel value will be stripped off the color values and converted to the grayscale. The pixel will have a value of 1byte per pixel which means any value between 0-255 is used. The grayscale image is then processed to find out possible number plate area using edge detection technique.

Detecting number plate using contours

After edge detection, all high intensity pixels were scanned from left to right in the X axis and top to bottom for Y axis. The concentration of white pixels will give a fair knowledge of where the more number of edges were available in x-axis. This will give the approximate starting location of x-axis. Concentration of white pixels in y-axis is also scanned to get the approximate location of the y-axis. Due to the variable text or images available in the vehicle, we may have more than one region of interest of the number plate.

Transforming Number Plate Image to Text:

The region of interest for the number plate area was detected from the previous module. We will try to fit the number plate area and ratio of width and height. If the possible number plates region is identified, the image is cropped for the dimension of the detected number plate. Then the convolution neural network (CNN) algorithm is performed. The number plate image is then processed.