融合卷积与自注意力机制的可见光-红外目标检测OA

In order to improve the detection performance of infrared and visible light target detection methods in complex environments such as uneven lighting,overlapping objects,and diverse object scales,a multimodal object detection model based on convolution and self-attention mechanisms is designed by combining convolutional neural networks with visual attention networks.The model adopts a lightweight network design concept to construct a multi-branch feature extraction structure based on the characteristics of infrared and visible light images.The visible light branch uses group convolution units as the core to progressively extract image texture details feature from a local perspective and the infrared branch utilizes an efficient multi-head self-attention module based on windows to extract significant image features from a global perspective.For different modal features,a cross-modal information fusion structure is designed to achieve complementarity between infrared and visible light features through key feature enhancement and dynamic adjustment of fusion weights.Meanwhile,to enhance the scale invariance of the model,a multi-scale feature filtering fusion structure is introduced to alleviate information conflicts between features of different dimensions from both channel and spatial location perspectives.Experiments on standard datasets show that compared with current similar methods,this method achives the detection accuracy of 85.3%in complex scenes,and it also has higher generalization ability and robustness.