The number of cotton seedlings can greatly contribute to the key seedling information, such as sowing survival rate and emergence rate. In this study, a cotton seedling counting algorithm was proposed using feature fusion. Five modules consisted of VGG basic, attention, feature fusion, and deredundancy normalization module. Among them, the VGG basic module was used the first four convolution layers of VGG-16, further removing the fifth convolution layer of VGG-16 and the subsequent pooling layer and full connection layer, in order to reduce the amount of calculation and model complexity. CBAM module included the Channel Attention Module (CAM) and Spatial Attention Module (SAM) to realize dual attention on the channel and spatial dimensions of features. The feature fusion module was used to fuse the features enhanced by the attention module and the first three layers of features extracted by the VGG basic module in three stages, thereby speeding up the learning speed of the model. Finally, the redundant information of the fused feature map was cleaned by the operation of deredundancy. The local counts of the single-channel feature map were redistributed by the operation of normalization. The algorithm flow was as follows. Firstly, the input image was used to generate a feature map through the VGG basic module. Then, the attention module was utilized to enhance the global information of the feature image through the channel and spatial dimensions. The feature fusion module was to fuse the enhanced features with the features in the basic module, and finally output the counting data after the removal of the redundant information and normalization. In addition, the field cotton seedling images were collected from 2017 to 2018. A new Cotton Seedling Counting dataset (CSC) was established using the images. A total of 212 572 seedlings were accurately and manually labeled in the dataset. Several typical target counting methods were compared on the CSC dataset, including MCNN, CSRNet, TasselNet, and MobileCount. Experiments showed that the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) of the counting were 63.46 and 81.33, respectively, indicating the lowest errors than that of the above four methods, while the average error decreased by 48.8% and 45.3%, respectively, indicating the effectiveness of the counting algorithm. Moreover, the test set was divided into two groups, where the midday group performed better in the light conditions, totaling 52 images. The morning and evening groups presented relatively weak light conditions, with a total of 47. Although photographed in different time periods and different light conditions, the performance of the model remained constant in the strong robustness and stability. when estimating the number of cotton seedlings in the actual cotton seedling images, the counting was still accurately estimated the number of seedlings under the interference of strong shadow noise, indicating the strong robustness. In addition, the Dropout feature fusion, attention module CBAM and normalization operation were verified by ablation experiments, indicating a positive impact on the model. Finally, image-based seedling counting can be widely expected to realize the scientific management of farmland. Anyway, the algorithm can be further optimized to explore the possibility of deploying on the real platform [ABSTRACT FROM AUTHOR]