D-GPH: Dynamic Graph-Prior Hybrid Detector with Uncertainty-Aware Refinement

Abstract

As a classic paradigm of two-stage target detection, Mask R-CNN achieves robust instance awareness through regional proposal and feature alignment. In its structure, the backbone network is responsible for extracting multi-scale features from the image layer by layer, and usually uses convolutional networks such as RESNET, which is difficult to use the global context prior to suppress the background noise in the degraded environment; Neck network (neck) relies on fixed multi-scale fusion scheme, lacks the ability of adaptive expression for different image contents, and is limited by the local receptive field of convolution operator, so its global modeling is easily disturbed by uncorrelated long-range targets; The detection head carries out category determination and boundary box regression on the candidate region (ROI), and usually uses the same set of forward and loss for difficult samples and easy samples. It is not enough to distinguish the fuzzy target with low confidence and small area, which is prone to missed detection and false detection. To solve these problems, this paper proposes a dynamic graph prior hybrid detection framework and named it D-GPH. Based on the global context prior generated by the backbone, the framework innovatively introduces the dynamic prior router (DPR), which generates adaptive prior injection strength for each layer according to the current image content, and realizes dynamic modulation by layer and graph before FPN fusion. In the neck design, this paper constructs Manhattan constraint graph attention module (MC-GAT), which explicitly suppresses non local noise by introducing spatial distance penalty, and cooperatively extracts global topology and local texture details with heterogeneous fusion layer. In the detection stage, this paper designed the uncertain perceptual refining head (UAR head), established the quality evaluation mechanism by using the exponential moving average (EMA) of varifocal loss and IOU, and started the secondary refining prediction based on channel recalibration for difficult samples, so as to significantly improve the positioning accuracy of complex targets. The experimental results show that the performance of D-GPH on MS coco dataset is better than Mask R-CNN benchmark, especially in the detection of fuzzy targets and difficult samples.