Embedded License Plate Recognition System Based on Blackfin561

With the development of science and technology and economy, the number of cars is gradually increasing, and traffic management problems are becoming increasingly severe. How to establish a scientific and effective traffic management system is the focus of traffic management. With the rise of science and technology, intelligent transportation system (TTS) is an inevitable trend in the development of future traffic management and a cutting-edge topic in electronic information technology. License Plate Recognition System (LPRS) is an important part of intelligent transportation system and is of great significance in automatic management, information statistics, vehicle tracking and other aspects. Although foreign countries have achieved good results in LPRS, China still needs to conduct in-depth research and implementation of many related technologies due to its late start.

The LPRS based on Blackfin561 replaces the role of manual license plate recognition, which can save a lot of manpower and material resources. The LPRS discussed in this article can be used at highway toll gates and community gates to achieve functions such as traffic flow statistics and vehicle tracking.

1 Overall design and architecture of the system

The embedded license plate recognition system is an independent system that can complete license plate recognition and information processing. The whole system includes video acquisition module, image processing module and communication module. The system software design includes six parts: road image acquisition, image preprocessing, license plate positioning, character segmentation, character recognition and result sending.

1.1 Embedded License Plate Recognition System Based on DSP

Due to its powerful data processing capability, DSP occupies an important position in the field of embedded image processing, especially dual-core DSP, which is rapidly rising in video processing applications. The embedded license plate recognition system based on DSP generally adopts a host and DSP distributed processing structure, which has the advantages of low price, good reliability and easy upgrade. The general structure of the system is mainly divided into three parts: image acquisition, DSP processing and PC data management. The system structure is shown in Figure 1.

The image acquisition module includes the video input part and the image extraction part. Its function is to collect brightness signals through the CCD camera, output analog composite video electrical signals, and decode the required digital image signals through the video decoding chip. The DSP processing module is mainly composed of the DSP processor and its peripheral circuits, external memory, and the interface circuit with the host. Among them, the external memory is used to store a large amount of image data for the DSP processor to call. Commonly used communication interfaces with the host include UART, PCI, USB, and Ethernet. Since the system usually needs to consider its own scalability, long-distance communication between the terminal and the host, and the difficulty of wiring, Ethernet is more commonly used, but the disadvantage is that on the DSP without integrated Ethernet function, an external chip is required, which increases the cost of the system. The PC data management module varies depending on the purpose of the system application, mainly including the human-computer interaction interface and the database and the corresponding processing mechanism.

1.2 System Hardware Architecture

The hardware platform architecture of license plate recognition based on Blackfin561 is shown in Figure 2, which mainly includes ADI's Blackfin561 processor, ADV7183A video decoding chip, Flash, SDR AM and DM9000. Blackfin561 resets and initializes ADV7183A and configures the working mode through the I2C bus. The video signal is collected by the parallel external interface (PPI) of Blackfin561 and stored in the external SDRAM. Blackfin561 calculates the image data and transmits the output recognition result to the host computer through DM9000 for subsequent processing.

1.3 System Software Design

The software flow of the license plate recognition system is shown in Figure 3, which mainly includes modules such as image acquisition, image preprocessing, license plate positioning, character segmentation and communication with the host computer.

1.3.1 Image acquisition

After the system is powered on, the DSP configures the registers of ADV7183A through the I2C bus, decodes the analog video signal collected by the camera, and outputs a digital video signal in the format of YCrCb in 4:2:2. Since the algorithm only processes the brightness signal, the 2-Byte sampling method is adopted through the two-dimensional DMA to extract only the brightness signal of the Y channel and store it in the SDRAM. At the same time, the ping-pong storage mechanism is used in the SDRAM to store two consecutive frames of images, so that the system can meet the requirements of real-time processing.

1.3.2 Image Preprocessing

Since the collected images may be polluted by noise due to factors such as weather, license plate cleanliness and light, resulting in a decrease in image quality, this will cause trouble for subsequent recognition work. Therefore, it is necessary to perform necessary preprocessing on the collected original images to improve the signal-to-noise ratio of the image and make the grayscale value and contrast reach an ideal state. Since the algorithm for denoising will blur the image edge to varying degrees, which is not conducive to the subsequent license plate positioning, the denoising work is placed after the license plate positioning. The preprocessing here mainly adjusts the grayscale range and contrast of the image through the histogram equalization method to achieve the purpose of image enhancement. The contrast effect before and after processing is shown in Figures 4 and 5.

1.3.3 License Plate Location

According to the relative position of the road and the camera, the speed range, and the road width information, the range of the license plate in the camera coordinate system can be determined in advance, which can greatly reduce the amount of calculation. In the predetermined license plate detection area, the image is binarized by the large rate method, and then the opening and closing operations in the morphological operation are performed to obtain the image shown in Figure 6. The image is projected vertically to find the row with the most white pixels, and then the boundaries are searched on both sides with this row as the center to find the upper and lower edges of the license plate. The image with the ordinate within the upper and lower edges is horizontally projected to find continuous white areas to determine the left and right boundaries. When searching for boundaries twice, the area of ​​the license plate must be determined in combination with the prior knowledge of the size of the license plate. If the size requirements are not met, the area found this time is excluded and searched again.

1.3.4 Character Segmentation

The original brightness data of the license plate area is extracted according to the boundary information of the license plate, and the license plate area is binarized using the large-scale method. This can eliminate the influence of other parts of the image on the grayscale range of the license plate. Then the binarized image is horizontally projected, and the license plate characters are segmented according to the width of the license plate characters. Each segmented character is then vertically projected, and the upper and lower black edges are removed according to the height size of the license plate characters to obtain the character image to be recognized, as shown in Figure 7.

1.3.5 Character Recognition

Character recognition uses a discrete Hopfield neural network. It is a recurrent neural network, and the output of the network will be fed back to the input, resulting in a constantly changing state. If the network is a stable network that can converge, the changes produced by this feedback and iterative calculation process will become smaller and smaller. Once a stable equilibrium state is reached, the network will output a stable constant value. The key is to determine its weight coefficient under stable conditions.

First, the standard characters of the license plate are discretized into a 16×32 binary image, that is, there are 512 neurons. In this way, the maximum number of samples that the network can remember is 512×0.15=76.8. According to the characteristics of the license plate, two networks are established: one is the network of license plate Chinese characters, including 36 Chinese characters; the other is the network of numbers and letters, including 36 elements.

It can be seen that neither network generates false samples. The standard character data is simulated by Matlab to obtain the stable weight coefficients of the network, and then the parameters are extracted for DSP calculation.

In DSP, the segmented character image is normalized to make the character image consistent with the size of the standard image, and then input into the network for calculation according to formula (1). Among them, Wij and θi are weight coefficients obtained by Matlab simulation, and x is the image data. Through cyclic iterative calculation, according to formula (2), when the energy function of the network reaches stability, the recognized character is obtained. The recognized character is subtracted from the standard character, and the difference is accumulated. The character with the smallest difference accumulation is the recognized result. Since the Hopfield network has the function of associative memory, it has strong anti-interference ability and has good recognition effect on noise, stroke disconnection, stroke adhesion and other phenomena, as shown in Figure 8.

1.3.6 Communication with the host computer

After the recognition is completed, the recognition results are sent to the PC in ASCII format through DM9000 for subsequent processing.

2 Conclusion

The key technologies of the license plate recognition system are license plate location, character segmentation and character recognition. This paper implements the three important parts in combination with the hardware platform. The experiment proves that the embedded system with Blanc fin561 dual-core DSP as the core and the Hopfield neural network can better complete the license plate recognition work, and can improve the speed and recognition rate comprehensively. The overall structure based on embedded system also makes the expansion of the system simple and has high practical application value.