For the High Energy Photon Source (HEPS) High-Performance Pixel Array Detector (HEPS-BPIX4), the HEPS-BPIX4 (High Energy Photon Source-Beijing PIXel4) DAQ data acquisition system must meet high real-time performance requirements. Online compression of image data can significantly reduce the pressure on subsequent data transmission and storage.

This study aims to overcome the limitations of traditional compression algorithms in terms of compression ratio and real-time performance by proposing an online image compression method based on deep learning object detection.

The end-to-end object detection model YOLOv10 was trained on an experimental dataset, and its training performance was tested and evaluated to ensure the model achieved the expected level of accuracy. Subsequently, the model's performance and effectiveness in data compression were tested and analyzed on the Intel Xeon 8462Y+ CPU and the NVIDIA A40 GPU. Finally, deployment of the model was optimized within the HEPS-BPIX4 DAQ framework under the multi-threaded scenario, and its practical performance was comparatively evaluated across different GPU platforms.

Experimental evaluations indicate that the proposed method achieves an average compression ratio of 5.88 for online image data. Furthermore, an efficient deployment strategy is devised and validated, achieving a compression processing rate in the GB?s^-1 range under single-threaded operation. Building upon this, a multi-threaded deployment framework for the HEPS-BPIX4 DAQ system is developed to fulfill more demanding compression performance requirements.

This research presents a novel approach to mitigate the processing burden imposed by high-bandwidth image data in the HEPS-BPIX4 DAQ system.