Study of Mask R-CNN in Target Detection Based on Improved Feature Pyramid Networks

In order to improve the effect of multi-scale object detection, this paper proposes an enhanced feature pyramid network (Enhanced FPN). On the basis of Mask R-CNN, the network introduces enhanced convolution block (Enhanced Con-volution Block) in the Neck part, which enhances the ability of feature extraction and fusion by combining conventional convolution, expansion convolution, deep separable convolution and channel attention mechanism (SE module). Enhanced FPN Contains an underlying feature extraction network and a multi-level feature pyramid to realize the fusion of multiscale features through top-down and bottom-up paths. During the fusion process, features of each scale are given adaptive learning weights to fully play the role of features at different scales. In addition, a new loss assessment mechanism (LossEvaIHook) is designed to dynamically evaluate model performance during training. The experimental results on the COCO 2017 dataset show that the average accuracy (AP) of this method reaches 52.5 %, which is significantly better than the benchmark model, and verifies the effectiveness and advantages of the proposed method.