Target detection method based on wide-area receptive field space attention

The invention discloses a target detection method based on wide-area receptive field space attention. The method comprises the following steps: preparing an image data set for training and testing; constructing a target detection network based on wide-area receptive field space attention, wherein the target detection network comprises four parts of Backbone, Neck, Head and MSA; and performing feature extraction on the test set image by using the trained network. According to the method, the pixel-level feature information is captured from the angle of a wide-area receptive field, meanwhile, mutual intersection between different feature information is considered, and the feature extraction effect is greatly improved under the condition that the parameter quantity and the calculation quantity are not obviously increased. 本发明公开一种基于广域感受野空间注意力的目标检测方法，包括：准备训练、测试用的图像数据集；构建基于广域感受野空间注意力的目标检测网络，包括Backbone、Neck、Head和MSA四部分；使用训练好的网络对测试集图像进行特征提取。本发明从广域感受野的角度捕获像素级特征信息，同时考虑了不同特征信息之间的相互交叉，在不显着增加参数量和计算量的情况下，大大提高了特征提取的效果。