Construction of a deep learning model for optical imaging target recognition under complex lighting conditions

Engineering notes

The study demonstrates that the deep model integrating illumination normalization, multi-domain adaptation, attention mechanism, and contrastive learning can significantly improve the comprehensive performance of the optical imaging target recognition system in complex lighting scenarios, providing a feasible idea for subsequent algorithm design and engineering deployment.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

In applications such as intelligent monitoring, autonomous driving, and unmanned aerial vehicle vision, complex lighting has become a significant factor affecting the accuracy and stability of optical imaging target recognition. This paper constructs a set of illumination-robust deep learning detection model. In terms of methods, a lightweight convolutional backbone and feature pyramid are used as the framework, and a light-weight illumination normalization and multi-domain adaptive enhancement network based on local statistics is introduced at the front end. In the middle, multi-scale illumination-invariant feature extraction is achieved through channel and spatial joint attention. During the training stage, complex lighting data augmentation and contrastive learning loss are fused, and the feature consistency of the same target under different lighting views is explicitly constrained. Experiments are conducted based on a dataset of 12,000 images covering uniform illumination, strong backlight, low illumination, and high dynamic range scenarios. The results show that: while the overall mAP is improved by approximately 3.8 percentage points, the mAP in strong backlight and low illumination scenarios is increased by approximately 13.7 and 15.3 percentage points respectively compared to the baseline, and the mAP fluctuation between illumination scenes converges from 18.1 percentage points to approximately 6.4 percentage points. The average inference latency is approximately 32.4 ms, and the FPS remains above 30 frames per second. The study demonstrates that the deep model integrating illumination normalization, multi-domain adaptation, attention mechanism, and contrastive learning can significantly improve the comprehensive performance of the optical imaging target recognition system in complex lighting scenarios, providing a feasible idea for subsequent algorithm design and engineering deployment.

Links and sources

Comments