Enhancing Zero-Shot Image Recognition in Vision-Language Models through Human-like Concept Guidance