Research And Application Of Lightweight Shadow Detection Algorithms Based On MobileNetV3

Shadow detection has always been a fundamental and challenging problem in the field of computer vision.In recent years,with the development of computer hardware,many traditional methods for shadow detection have been replaced by deep learning shadow detection methods.The shadow detection method has become a new research hotspot,which promotes the development of shadow detection field with high detection accuracy and strong generalization ability.It is aimed at the limitation of the number of layers of the deep learning network and the insufficient hardware performance of the mounted device when the shadow detection task is directly performed on a small mobile device.With the characteristics of less parameters and less computation,the lightweight network is more suitable for mobile shadow detection tasks with limited storage space and power consumption.It has excellent performance and is sought after by academia and industry.The application research of MobileNetV3 in the field of shadow detection has also become a hot topic.In order to improve the shadow detection effect of the model under the premise of effectively controlling the network scale,so that the lightweight network can be effectively used for mobile shadow detection tasks,this paper makes further improvements based on MobileNetV3,and proposes two different shadow detection network models.Implement lightweight shadow detection tasks.The research in this paper is as follows:(1)A dual-stream feature extraction shadow detection network based on MobileNetV3 is proposed.The extraction of complex shadow features is enhanced by designing a feature flow of detail image feature extraction and a feature flow of global image semantic extraction to form dual-stream features,and information fusion of shadow features in the two feature streams is performed to comprehensively utilize shadow detail feature information and global shadow location information.Their respective advantages strengthen the feature extraction capability of lightweight networks.Experiments show that compared with the MobileNetV3 network,the network can greatly improve the quality of shadow detection while keeping the network parameters low.(2)A lightweight shadow detection network fused with dilated convolutions is proposed.By analyzing and improving the feature extraction process of MobileNetV3,the network designs a dilated convolution bottleneck structure to enhance the feature extraction capability of the network,comprehensively utilizes multi-scale shadow context information,and at the same time,constructs spatial information compensation structure in key feature layers,reduce the loss of spatial semantic information of shadow image in the process of downsampling of shadow feature map,and finally restore to the input image size through layer-by-layer feature merging and upsampling.It has been proved by experiments that compared with MobileNetV3 network,this network has fewer networks.The amount of parameters and higher shadow detection accuracy can be easily deployed on resource-constrained mobile devices.(3)A lightweight shadow detection system is designed and implemented.The system is deployed using the models trained by the above two shadow detection models.Users can select shadow images to be detected,train and predict through two different lightweight shadow detection networks,and output shadow detection results to provide users with reliable mobile terminals.Shadow detection service.