基于先验框设计(anchor-based)的多类目标检测算法存在超参数多、泛化能力差、正负样本不平衡的问题。针对这些问题,提出一种基于改进无锚(anchor-free)方法的目标检测算法。首先,针对传统算法在多类目标检测任务中难以获得鲁棒的特征表达的问题,构建基于上下文结合的自校准双重注意力模块,通过混合空洞卷积组获取多感受野信息;然后以低维空间嵌入的方式进行自校准获取上下文空间信息;最后将空间信息与通道信息结合,增强算法特征表达能力。针对在多类目标检测任务中由于目标尺度变化大、外观不规则而易引入背景噪声的问题,利用改进的可变卷积,对目标进行自适应采样。在目标检测数据集MSCOCO上的实验结果表明,所提算法能有效提升目标检测精度,优于对比检测算法。

To prevent numerous hyperparameters and to overcome poor generalization ability and imbalance between positive and negative samples in anchor-based multiclass object detection algorithms, an object detection algorithm based on an improved anchor-free method is proposed herein. To address the difficulty faced by traditional algorithms in obtaining robust feature representations in multiclass object detection tasks, a self-calibration dual-attention module based on contextual combination is first constructed herein. It obtains the multireceptive field information through a mixed dilated convolution group. Then, a low-dimensional spatial embedding method is self-calibrated to obtain the contextual spatial information. Finally, the spatial information and channel information are combined to enhance the feature representation ability of the proposed algorithm. To prevent the usual introduction of background noise owing to large changes of object scale and irregular appearance in multiclass object detection tasks, the improved deformable convolution is used to adaptively sample the target position. Experimental results obtained using the large multiclass object detection data set MSCOCO show that the proposed algorithm can effectively improve the detection accuracy of multiclass object and outperforms the existing detection algorithms.