面向暗光场景的目标偏振/可见光融合检测方法

Target polarization/visible light fusion detection method for dark light scenes

摘要:

为解决偏振暗光场景下常见目标识别结果准确性不高的问题，提出了基于卷积神经网络的偏振度图像与可见光图像融合算法，设计了新的损失函数以形成无监督学习过程，引入拉普拉斯算子提高融合图像的质量，最终将被测目标的偏振信息与可见光信息有效结合；提出了基于改进的YOLOv5算法对融合后的目标进行目标检测，在网络框架中加入CA注意力机制，将通道注意力机制与空间注意力机制相结合。在自制的数据集上对提出的网络进行训练测试，结果表明，融合图像在主客观上都达到了较好的视觉效果，将改进的YOLOv5算法相比最优的YOLOv5s模型，精确率和召回率分别达到了89.3%和82.5%，均值平均精度分别提高了2.6%和1.8%。

Abstract:

In order to solve the problem of low accuracy of common target recognition results in polarized dark light scenes, a fusion algorithm based on convolutional neural network of polarization degree image and visible light image is proposed, a new loss function is designed to form an unsupervised learning process, the Laplace operator is introduced to improve the quality of fused image, and finally the polarization information of the target to be measured is effectively combined with the visible light information; and a fused target detection algorithm is proposed based on the improved YOLOv5 algorithm for target detection of the fused target, adding CA attention mechanism to the network framework, combining the channel attention mechanism with the spatial attention mechanism. The proposed network is trained and tested on a homemade dataset, and the results show that the fused image achieves better visual effects subjectively and objectively, comparing the improved YOLOv5 algorithm to the optimal YOLOv5s model, the precision and recall reach 89.3% and 82.5%, respectively, and the mean average precision is improved by 2.6% and 1.8%, respectively.