By reason of its wide range of applications, object counting is increasingly recognized as a prominent branch of computer vision. Moreover, object counting has an active role in poultry statistics and fruit yield prediction in the agricultural field. However, in a dense target counting study, annotating a large-scale object counting dataset at the pixel level is labor-intensive and time-consuming. To alleviate dependence on labeled data and take advantage of unlabeled data, we take dense crowd counts as the subject of our study and propose a novel mutual consistency learning model (MCL) for semi-supervised crowd counting. This model includes two networks with the same structure but different initialization. By imposing consistency constraints, the two networks can be encouraged to interact effectively in high-density regions. We further introduce an observation-based random image recombination (RIR) approach, which splits an image into several equal patches, and then recombines these patches into an image with the same shape as the initial image without changing the population. The RIR disrupts the scene spatial order in the image, enhancing the model's ability to adapt to a deceptive environment and improving the ability to retrieve crowd characteristics. The presented method delivers comparable performance on three challenging benchmarks compared with other semi-supervised methods. The outcomes also indicated that the suggested method could achieve excellent results and provide an algorithmic basis for future applications in agriculture, such as sheep counting, duck counting, fruit counting, etc.