Research On High-quality Multimodal Style Transfer Method For Road Scenes

Image style transfer is an important research issue in the field of computer vision and pattern recognition.Scene modeling based on road scene image has important applications in off-line test of unmanned vehicles and intelligent robot evaluation.However,road scene modeling needs to collect large-scale image data with high-precision,which is time-consuming and inefficient.Therefore,the image style transfer technology can be used to generate the road scene style image under different lighting and weather conditions.Based on the above problems,firstly,this thesis proposes a high-quality single mode style transfer network DFGAN for road scenes.The network is based on the dual branch structure,one branch extracts the detail features of the image through the super-resolution task,and the other branch generates the transferred features based on the encoding and decoding structure.The network also contains filter module,which can effectively suppress the style information contained in the detail features,and avoid the influence of different styles superimposed in the generated results.Compared with the existing methods,the proposed network can generate images with more texture and details.Furthermore,in order to enhance the richness and diversity of generated results,this thesis proposes a high-quality multimodal style transfer network CLGAN for road scenes.Based on the assumption that image can be encoded into shared content feature and exclusive style feature,the original generator in DFGAN is improved to an adaptive generator with separable encoding and decoding process.Meanwhile,the network introduces the idea of contrastive learning in training,which replaces the cycle consistency constraint,so as to improve the training efficiency.In actual test environments,the network can generate multimodal high-quality images with the style of the target domain based on the image detail features,content features and random sampling style coding in the style space of the target domain.Experiments on public datasets show that the proposed methods can generate road scene images with high-quality,which proves the effectiveness of the proposed methods.Compared with the benchmark method,the experimental results show that the proposed method CLAGN can generate multimodal transferred image data with high-quality for large-scale road scenes.