基于特征融合的可见光与红外图像目标检测

为提升目标检测任务在复杂环境下的识别效果，提出了一种基于特征融合的红外与可见光目标检测方法。该方法首先采用并列的卷积神经网络分别提取红外和可见光特征信息，并利用通道和空间注意力机制提升有效特征的权重；其次，为充分利用红外和可见光特征进行信息互补，设计了特征自适应融合结构，以自主学习方式将红外与可见光特征以最优方式加权融合；最后，针对不同尺度目标，通过交替采样方式充分融合深层和浅层特征，保障各维度目标检测效果。通过实验表明，所提方法可以充分利用并融合不同模式、尺度的目标特征信息，实现目标准确识别及定位。同时，在实际电网设备检测中，该方法也体现出较优的鲁棒性和泛化性。

Object detection in visible light and infrared images based on feature fusion

In order to improve the recognition of object detection task in complex environment,an infrared and visible target detection method based on feature fusion is proposed.Firstly,the parallel convolution neural network is used to extract the infrared and visible feature information respectively,and the channel and spatial attention mechanism are used to enhance the weight of effective features.Secondly,in order to make full use of the infrared and visible features for information complementarity,a feature adaptive fusion structure is designed to autonomously learn to weight infrared and visible features together in an optimal way.Finally,for different scales of targets,the deep and shallow features are fully integrated by alternating sampling to ensure the object detection effect of each dimension.The experiments show that the proposed method can make full use of and integrate the target feature information of different modes and scales to achieve accurate target recognition and localization.At the same time,this method also shows better robustness and generalization in the actual power grid equipment detection.