Object Detection Method in Foggy Images with Dynamic Convolution Kernels and Crossdimensional Attention

To address the issues of difficult extraction of effective image features and poor performance in small object detection caused by fog occlusion, an object detection method in foggy images with dynamic convolution kernels and crossdimensional attention is proposed. First, a dynamic convolution kernel module is designed, which multiplies the local attention scores with the convolution kernel parameters at corresponding positions, enabling adaptive adjustment of the convolution kernel parameters based on local features and enhancing the feature extraction capability of the kernels. Second, cross-dimensional attention is constructed to achieve multi-dimensional feature complementarity through the interaction and fusion of spatial and channel dimensions. Additionally, a dedicated small object detection head is introduced, which takes the cross-dimensional attention as input to improve the accuracy of small object detection. Experimental results show that the improved algorithm achieves mAP@0.5 scores of 73.1% and 52.5% on the public datasets RTTS and Foggy Cityscapes, respectively, representing improvements of 3.8 and 5.6 percentage points over the baseline algorithm.