DH Detector: One-stage detection model with double-head detector and efficient convolution

This article explores advanced techniques to build a network model, including the use of a double-head(DH) detector: one section of which is responsible for performing classification and the other for object prediction. This is followed by information stitching, which allows the model to learn more appropriate parameters. Subsequently, a CSP structure incorporating an attention mechanism is used to improve the feature extraction capabilities of the model. Finally, dynamic convolution is added to enhance the feature representation. Our model is trained from scratch on the COCO Dataset without taking a pre-trained model, the results of which are listed below. Based on 300 epochs of training, the mAP values for the four versions of DH Detector (s, m, l, x, the network model becomes larger in turn) are 40.8, 46.1, 48, and 49.8(100 epoch), respectively. The experiment results show that this method outperforms other competing methods in terms of both quantity and quality. In addition, the mAP of DH-s and DH-l reached 43.0 and 50.1 at an input image size of 960. A comparison of DH Detector with other algorithms is provided in Figure 1.