“Large scale visual language models have made progress in the field of object counting, but face two major challenges: class semantic misalignment and decoder architecture limitations. Experts propose the Cross Branch Collaborative Alignment Network (CANet), which adopts a dual branch decoder architecture and visual text category alignment loss to effectively solve the above problems and achieve excellent performance on multiple benchmark datasets, providing new ideas for improving counting robustness in complex scenes.”