Implementation of Multiple DSP image processing and recognition systems for Marine Search and Rescue

Implementation of Multiple DSP image processing and recognition systems for Marine Search and Rescue

[Date:]

Source: China Power Grid Author: Bi Wenyu Zuo wenxu Bing, School of opto-electronics, Yantai University

[Font:Large Medium Small]

　

Introduction

Maritime search and rescue services are an important part of the national emergency rescue system and an important guarantee for national economic development. The adoption of advanced search systems is one of the important means to improve the effectiveness of search and rescue operations. Generally, the search system uses a radar or photoelectric imaging system. Because the optical image sensor has a high resolution, the observation distance can reach 20 ~ 40 km, but its disadvantage is that it depends heavily on good weather and sunshine conditions. However, infrared image sensors can penetrate smoke, smog, haze, and snow to make up for this deficiency. Therefore, they use a photoelectric search system consisting of infrared, visible light imaging, and DSP image processing systems.

According to the nature of marine search and rescue, the system should have the following basic capabilities:

· High suspicious target detection capability;

· High anti-interference capability of the Marine background;

· Has certain target recognition and tracking capabilities.

Based on the above considerations, this paper designs an image processing system consisting of two pieces of and eight pieces of T.

System Metric requirements

· Visible video input: CCIR/EIA or PAL/NTSC video signal

· Infrared video input: CCIR/EIA

· Video output: VGA or standard video output

· Video adconversion precision: 12bit

· DA conversion precision of video display: 10bit

· Image processing frame rate: greater than 30Hz

· The smallest prime number of the "target" can be detected: 3 × 3

System Solution

System diagram 1.

Video Acquisition is achieved using, and multiple 64-16-HPI interfaces are connected through the XINTF interface to achieve image data transmission. Each 6416 adopts the flow mode, and outputs the processed image data through the EMIFB interface. After the display circuit, the monitor displays the output image.

The system adopts an expandable structure and determines the number of 6416 image processing units as needed. Generally, four or eight 6416 image processing units are used.

The system software filters image data and performs edge detection to detect and identify suspicious targets.

When inputting a single video, you can select four or eight 6416 image processing units to form an image processing machine.

Figure 2 shows the sequence of double video input. Each video channel has four 6416 image processing units, each of which adopts a flow mode. Each video channel processes a maximum of four frames at a time, the two groups of processing results are output to the image display module for processing.

System Hardware Design

The hardware system consists of three parts: Video Acquisition Circuit, image processing circuit, and display circuit. Figure 3 shows the overall hardware design.

Video Acquisition Circuit

Figure 4 shows the principle of Video signal acquisition based on 2812-DSP, including the video preprocessing module and 2812 module. The video preprocessing module consists of Y/C separation, level clamp, synchronous separation, and amplitude adjustment processing circuit. On the 2812-DSP, A/D collects video signals at A speed of MB, reaching the extreme sampling rate (the sampling interval is 80 ns ).

The principle of Y/C separation, video clamp, and synchronous separation circuit is shown in Figure 5.

TMS320C6416T sub-module

This module is the core module of the processing part of the system. 6416 sub-modules are designed based on the design philosophy of strong versatility, clear and simple interfaces, resource maximization, and multi-6416 system construction, as shown in figure 6. The EMIFA interface is used to expand two pieces of 4 MB × 32 bit SDRAM, which can read and write 64 bit data at a time.

Image Transmission Interface Design

The image data is transmitted through the 2812-DSP-XINTF (16bit)/6416-DSP-HPI (32bit) interface, and the 2812-DSP output results are spliced into 32bit using a single CPLD. Optimizing the 2812-DSP-XINTF register can maximize the HPI interface transmission rate. The specific optimization values for the XINTF register are shown in table 1.

Image Display

Use a VGA Monitor to display image processing results. Standard SVGA interface signals include: line synchronous signals (VGA_Hs), field synchronous signals (VGA_Vs), and red, green, and blue analog signals. The timing synchronization signals required by VGA are generated by CPLD, And the analog signals required are implemented by the video D/A converter ADV7123. The Interface Circuit Diagram 7 is displayed.

The CPLD performs bus arbitration on the EMIFB bus of each 6416 image processing single metadata output interface (EMIFB) to implement time-based output of each 6416 unit image data.

Two pieces of SRAM using the "ping-pong access" method constitute an image data buffer. Each piece of SRAM stores one frame of image, which is controlled by CPLD.

System Software Design

The system software flowchart 8 is shown. There are three main parts: Image preprocessing, suspicious object extraction, and object recognition.

Image preprocessing

Image filtering (Multi-template composite filtering algorithm)

Compared with common filtering algorithms, a filtering algorithm is only effective for partial noise. It is difficult to maintain image clarity while effectively suppressing noise. The adoption of multi-template composite filtering algorithm can better solve this problem and lay the foundation for edge extraction.

Edge Detection (improved sobel operator)

Common Edge Detection Algorithms are seriously affected by sea surface ripple. In contrast, the sobel operator works better, but there are also situations where Edge Points are missed. Based on the two templates of the traditional sobel operator, this algorithm adds six templates and uses the following eight templates to calculate each pixel respectively. Then, the maximum value is used to replace the value of this pixel.

This improved sobel operator makes edge detection more accurate, but the processing of a frame of image requires a large amount of work. After the improved sobel operator is used for image edge detection, the image is binarization. The improved sobel operator not only detects all suspicious targets, but also has little impact on the waves.

Suspicious Target Extraction and tracking

To reduce the amount of computation data, use a secondary tag to extract suspicious targets:

· Use the labeling algorithm to pre-mark the number of objects and record their locations;

· Extract suspicious targets based on the template information provided by the upper computer;

· Perform secondary marking to only mark the suspicious targets that have been extracted;

· Tracking suspicious extracted targets.

Target recognition and tracking

When the extracted suspicious target (usually a small target with few prime numbers) reaches a certain prime number, the Hu Moment-changing feature is used to identify the suspicious target.

For digital images f (x, y), p + q moment (mpq) and center moment (μ pq) are defined

P and q are non-negative integers. The center moment of the (p + q) Normalization of the image is defined:

Use the center moment of the second and third-order normalization to find the seven Hu Moment groups:

A (x, y) is the correlation value of the moment at the position (x, y. Take the point corresponding to the maximum value of A as the matching point.

Immutations describe the statistical characteristics of images and meet the immutability of changes such as translation, scaling, and rotation. Therefore, they are widely used in image recognition and other fields. The disadvantage of this algorithm is the large computing volume. However, simply processing the image data of a local suspicious target area can greatly reduce the amount of data computing. Add and track the identified targets.

Conclusion

· A multi-DSP image processing and Recognition System for Marine Search and Rescue is successfully implemented;

· 2812 video data acquisition is successful;

· The scalability of system hardware enhances the versatility of the system;

· The actual application of software algorithms is remarkable.

The system (PCB Board shown in Figure 9) can also achieve Passive Ranging for search and rescue targets, dual-band image fusion of visible and long-wave infrared images, and other functions. In addition, the powerful processing capabilities and storage space of the system enable it to play a greater role in the digital image processing field.