InternDiffuseDet: Object Detection Method Combining Deformable Convolution and Diffusion Model

The paper focuses on the topic of object detection and aims to address issues such as missed detections, limited feature extraction capability, and low detection accuracy in complex scenes. Building upon DiffusionDet, a modified approach is proposed that combines deformable convolutions and diffusion models for object detection. The core idea is to increase the quantity and quality of feature maps before entering the detection head. This is achieved by introducing InternImage and DCNv3 deformable convolution operators into the backbone network, enhancing the receptive field and non-linear modeling capability of the model. An improved feature pyramid network (CS-FPN) based on selective weighting is proposed to enhance the intermediate FPN feature pyramids. Channel and spatial separations are achieved using depth-wise separable convolutions, with the traditional upsampling operation being replaced by the CARAFE operator to improve resolution and semantic information transfer. Following that, the SGE attention mechanism is employed to reassemble the feature maps, ensuring the preservation of hierarchical information during diffusion. Prior to entering the detection head, the DDIM diffusion operation is performed to obtain feature maps at different time steps, thereby augmenting the quantity of detection feature maps. Finally, the EIOU algorithm is introduced in target box matching and loss functions to handle position deviations and scale differences between target boxes. Experimental results on the COCO dataset and road detection dataset demonstrate that the improved model is 3.8 and 3.6 percentage points higher than the original model, respectively, in the same experimental settings. These results indicate the potential of the proposed method to enhance the accuracy and robustness of object detection, providing new insights and approaches for addressing object detection challenges in real-world scenarios. © 2024 Journal of Computer Engineering and Applications Beijing Co., Ltd.; Science Press. All rights reserved.