|
Firstly, the core processing process of current inversion algorithms is analyzed and a flowchart is provided, including four key steps: image reading, target tracking extraction, inversion calculation, and result storage. Using program instrumentation, the execution time of each step in the inversion calculation process is obtained. The image reading takes about 5.4ms, the target tracking extraction takes about 22.5ms, the inversion calculation takes about 1.0ms, and the result saving takes about 16.2ms. Secondly, we propose a "producer/consumer-producer/consumer" pattern of infrared image inversion. Each part can be executed synchronously by different threads or thread groups. We test the execution time of each step and divide them into three weakly coupled modules: producer, consumer-producer, and consumer. The producer corresponds to image reading; the consumer corresponds to result storage; since the target tracking extraction process takes much longer than the inversion calculation process and the two processes are closely related, we combine them into the consumer-producer, which is both a consumer (relative to upstream producer) and a producer (relative to downstream consumer). Thirdly, we determine the number of threads for producer, consumer-producer, and consumer. Considering that the target tracking extraction algorithm in the actual image inversion process requires continuous image frames, the consumer-producer can only use a single thread to execute, while image reading (producer) and result storage (consumer) consume relatively less time compared to the consumer-producer. Therefore, in this case, when producer, consumer-producer, and consumer are all implemented using one thread, we can achieve the optimal thread configuration. The paper presents a logical diagram and pseudocode implementation to show the process of three threads synchronizing and mutual exclusion to perform the infrared image inversion calculation. Finally, we implement the improved inversion pattern using C++ and test and compare it with the current infrared image inversion algorithm at the missile range. We used 4,000 actual measured infrared images of a certain aircraft, with a target size of approximately 30,000 pixels. We processed the images using two different methods four times, and calculated the average inversion computation time per infrared image. The results show that compared to the current infrared image inversion process, the improved pattern improves efficiency by about 48% and can meet the real-time inversion requirements at the missile range. This scheme and its design ideas and methods can be applied and promoted in various fields involving data acquisition and processing. |
ACCESS THE FULL ARTICLE
No SPIE Account? Create one
Image processing
Infrared radiation
Infrared imaging
Data processing
