| Object detection refers to the task of detecting specific objects in an image and locating them,which is a fundamental task in computer vision analysis and understanding.In recent years,significant progress has been made in image object detection with the development of deep learning techniques,and many excellent object detection models have emerged.Among them,the keypoint-based object detection method determines the category,location,and shape of objects in an image by detecting one or more keypoint information.It not only detects objects more accurately but also adapts better to changes in target scale,rotation,and other situations,with stronger robustness.Therefore,this paper aims to further improve the detection accuracy and performance of keypoint-based image object detection models from the perspectives of lightweight network models and improved loss functions.The main research contents of this paper are as follows:(1)To address the issues of high complexity and huge parameter count in keypoint-based object detection network models,this paper proposes a lightweight object detection method based on the Hourglass network.Instead of the original Hourglass network’s upsampling and downsampling blocks,this paper uses backward residual blocks based on depth-separable convolution.Meanwhile,a multi-scale feature fusion method is adopted to compensate for the loss of accuracy caused by lightweight design and reduce the semantic confusion caused by directly stacking features from different network layers.Comparative experiments were conducted on the PASCAL VOC public dataset and a mask image dataset.The results show that the proposed method achieves an average precision about 1% lower than the original Center Net method,but the parameter scale and floating-point computation are reduced to 4% and 5% of the original Center Net method,respectively.The detection speed is improved by nearly 3 times,and the model has good performance in balancing detection efficiency and accuracy.(2)To further address the issue of insufficient accuracy in object detectors with stacked Hourglass networks as the backbone,this paper proposes an improved cascade focal loss function.This loss function allows the shallow network in the Hourglass network to be the main identifier for target recognition tasks,while the deep network is responsible for identifying difficult and small samples,fully leveraging the recognition performance of different levels in the Hourglass network,and improving the accuracy and performance of keypoint-based image object detection.This paper applies the improved cascade focal loss function to three object detectors,Corner Net,Tri＿Center Net,and Single＿Center Net.Experimental results on the COCO dataset show that the proposed method improves the m AP values of the original detectors by 1.2%,0.3%,and 0.7%,respectively.The m AP values in small target detection also improved by 0.8%,0.6%,and 0.6% respectively.Among them,the m AP value of the Single＿Center Net detector reaches 46.5%,achieving the best detection accuracy in single-stage object detection models.This indicates that the proposed method can effectively improve model performance and object detection accuracy. |
