Abstract: The study was conducted on the detection of tomato fruit maturity and appearance quality based on deep learning methods using pink tomato as the experimental material. Two thousand and thirty-six tomato image data were collected and amplified to 5 316 through preprocessing. Then，the data was annotated and converted into files to construct an experimental dataset. The experiment improves the accuracy of SC-YOLOv5s by adding CA attention mechanism，replacing the Stem block structure，optimizing the detection layer scale based on recognition requirements，and replacing the K-means++ clustering algorithm to improve the model’s feature expression ability. By adding a fire module structure to SC-YOLOv5s for lightweight convolution and reducing the parameter count of the Bottleneck module，the SC-YOLOv5s-lite lightweight design is achieved，improving the detection speed of the model on hardware；Train and optimize the SC-YOLOv5s-lite model on the training set. The results showed that the memory usage of the SC-YOLOv5s-lite model was 7.73 M，with an accuracy rate of 89.04%，a recall rate of 83.35%，an average accuracy of 91.34%，and a detection time of 143 ms. Compared to YOLOv5s，the model parameter quantity is reduced by 54.57%，model size is compressed by 44.86%，with an average accuracy improvement of 3.98%，and the detection time is reduced by 20.99%. It has obvious advantages and is more suitable for hardware deployment.

Key words: tomato, maturity, appearance quality, detection, deep learning, computer vision