Abstract: Aiming at the problems of single shot multibox detector（SSD） network model with large parameters, poor detection of small targets and low detection accuracy of crops and weeds, a weed detection method based on multi-scale fusion module and feature enhancement was proposed. Firstly, MobileNet, a lightweight network, was used as the feature extraction network of SSD model to reduce the amount of model parameters and improve the speed of model feature extraction. And a multi-scale fusion module was designed to enhance the key information in the shallow feature map by channel attention mechanism, and then the receptive field was expanded by dilated convolution with different expansion rates. Finally, the two branches were fused, so that the shallow feature map used to detect small targets can contain rich semantic information while containing more detailed information of small targets. On this basis, the output six feature maps were feature enhanced by the channel attention mechanism to enhance the key features in the images and make the extracted features more directional, thus improving the detection accuracy of the model for crops and weeds. The experimental results showed that the weed detection model based on multi-scale fusion module and feature enhancement proposed can achieve an average detection accuracy of 88.84% in the image data set of sugar beet and weeds in the natural environment, which was 3.23 percentage points better than that of the standard SSD model, 57.09% less parameters, and 88.44% faster detection speed, while the model’s ability to detect small-scale crops and weeds, and leaf overlap were all improved.