Detecting and classifying apples based on their ripeness in natural environments is crucial. This paper proposes an apple maturity detection model utilizing GSEM_YOLOv5s. Initially, apple ripeness is categorized into fully ripe, medium ripe, and unripe, according to market requirements. To enhance model performance, Ghost Conv and Ghost Bottleneck techniques are incorporated to reduce parameter count while maintaining accuracy. Additionally, the SE attention module is integrated to improve key characteristic perception. The CIOU function is replaced by the EIOU function to expedite boundary regression. Furthermore, the MetaACON activation function is introduced to enhance information utilization. Compared to the original YOLOv5s model, the improved model increases detection accuracy for fully ripe, medium ripe, and unripe apples by 1%, 11%, and 8%, respectively, and reduces parameter count by approximately 15%.