Research on Multitask Deep Learning Network for Semantic Segmentation and Object Detection

Abstract

After analyzing methods of object detection under the existing deep learning framework, a multitask learning model (Fully Convolution Object Detection Network, FCDN) is proposed, which can realize complete end to end semantic segmentation and object detection through deep learning, without delimiting the default boxes. First, this paper analysis the reason why the current mainstream object detection network needs the default box delineated in advance; second, an object detection network with no delimited default box needed is proposed. It uses the semantic segmentation to detect all boundaries and key points of object at the pixel level, and then obtain prediction boxes by combining the category information of the semantic segmentation map. Finally, the feasibility of the method is verified on the VOC 2007 datasets, and compared with the performance of current mainstream object detection algorithm. Results show that the semantic segmentation and object detection can be realized at the same time by the new model. Trained by the same training sample, detection precision of FCDN is superior to that of classic detection models.