Dark Object Detection Framework Based On Semi-decoupling

Object detection on well-exposed images has attracted widespread attention and made breakthroughs in the fields of autonomous driving and video surveillance.However,the performance of general object detectors based on low-light environments has declined dramatically.The performance gap is mainly due to the lack of labeled dark object datasets in this field and the illumination sensitivity of the deep features mined by the general feature map extraction network.Some work for low-light object detection is based on the method of brightness enhancement to alleviate the feature degradation caused by the change of dark light data distribution,but the inevitable noise expansion in the process of brightness enhancement will affect the detection performance of downstream tasks.Therefore,most of the current dark light target detection work is devoted to aligning domain features and improving the robustness of model environmental conditions.A general,reasonable,and effective reduction of domain spacing is an urgent problem to be solved in the field of low light object detection.To solve the above problems,this paper proposes a framework for dual-task dark light target detection based on semi-decoupling.The main research contents are as follows :(1)We propose a general algorithm framework for low-light object detection,which effectively reduces the inter-domain feature distance.The low-light framework consists of a dual-task encoding-decoding network for image degradation and low-light object detection.The image degradation network uses the camera-based image processing operator to degrade the normal illumination image to alleviate the scarcity of low-light data sets,strengthen the supervision of the shared weight encoder,and provide high-quality features for the object detection network.(2)We propose a semi-decoupling scheme for the entanglement between tasks in the low-light frame to achieve fine-grained unwrapping of the features of the hierarchical encoder and improve the effectiveness of dual-task joint training.The experimental results show that the detection accuracy of the low-light domain is significantly improved after the low-light frame is embedded in the general object detector.(3)We design a semantic feature fusion network to alleviate the lack of shallow feature space information and the problem that the feature features of image degradation task learning pay attention to non-semantic regions.On the one hand,the network integrates long-distance information into rich spatial information and lacks advanced information.