CCSFNet: An efficient model for safflower maturity detection

Safflower is a medicinal crop with important economic value, but its flowering period is short and the flower clusters open in batches. Whether it can be harvested in time will directly affect the quality of safflower. In order to realize the intelligent detection of safflower maturity, this paper proposes a safflower maturity detection method based on CCSFNet. A central attention mechanism (CAM) is introduced to achieve attention superposition on key areas to enhance the extraction of key features of red flower images. To reduce the computational burden of the attention mechanism, we integrate depthwise separable convolutions and enhanced bottleneck units into a multi-branch residual design. By combining progressive feature reuse with efficient convolution operations, we achieve a lightweight feature extraction module, C3k2-e. During the model training phase, an improved genetic algorithm Fast GA is used for hyperparameter optimization, which significantly reduces the algorithm iteration time while maintaining the optimization effect. Compared with the current best detection models such as YOLOv11n, YOLOv10m, and RT-DETRv2s, CCSFNet achieves a better balance between detection accuracy and speed. The final optimized version maintains an AP of 97.39%, while the inference latency is only 1.58ms, which is an improvement of 2.06% and 0.06ms compared to YOLOv12, and an improvement of 1.66% and 3.44ms compared to RT-DETRv2s. Finally, it was successfully deployed on Jetson Orin Nano, providing strong technical support for automated safflower harvesting and quality control. The relevant code is open source at https://github.com/mpwmpw/CCSFNet