Crowd counting is a challenging task with important applications in public safety, traffic control and planning. Weakly supervised crowd counting models, in contrast to fully supervised models, do not require object-level annotations during training, which can reduce the cost of manual labeling. However, the lack of object-level constraints often leads to poor performance. To address this, we propose a weakly supervised crowd counting method that combines contrastive deep supervision and feature fusion. During training, we project intermediate feature maps from different stages and perform contrastive deep supervision, which strengthens the model's understanding of crowd features. The model also introduces feature fusion to allow the regression layer to fully utilize feature information from both deep and shallow layers for regression counting. Additionally, we introduce an attention mechanism in the forward process to enhance the weights of important crowd features. Experimental results show that our method achieves good performance on multiple crowd datasets, even better than some fully supervised methods.