Design and implementation of a compact fresh tea-leaf sorting system for integrated harvesting–sorting equipment in hilly mountainous tea garden

Mechanically harvested tea leaves show poor uniformity and cannot directly meet the raw material homogeneity required for famous tea processing. This study develops a compact fresh leaf sorting system for hilly mountainous tea garden that can be integrated into a cross ridge continuous harvesting machine. The system combines an electromagnetic vibrating screen with a two-stage differential speed conveying unit to reduce adhesion and accumulation of machine harvested fresh leaves. Tea-leaf views are then obtained via region segmentation guided by color-index features. On this basis, a lightweight classification network with two primary convolutional layers and multi branch depthwise separable convolutions is designed to recognize three grades: broken and single leaves, one bud and one leaf, and one bud and multiple leaves. A dynamic ROI window strategy that tracks and enlarges boundary leaves across frames is further proposed. Together with multithreaded ONNX Runtime inference and a multi-channel FPGA delayed air jet execution unit, this enables closed loop control from image recognition to air jet sorting. With about 3.7 M parameters and 0.04 G FLOPs, the network achieves a three grades classification accuracy of 96.36%. When the visible field of boundary leaves across frames is increased from 50% to 100%, the dynamic ROI improves recognition accuracy by approximately 42.3%. In a configuration that uses only a CPU, the system attains an online inference speed from 50 to 70 frames per second, an average sorting accuracy of about 82% for mixed feeding of the three grades, and fresh leaf uniformity that meets the grade 1 and grade 2 requirements of GB/T 18650–2008, with engineering-estimated single-channel throughput of approximately 26.7 kg/h. These results show that the proposed system provides a feasible technical route for integrated mechanized harvesting and sorting of famous tea in hilly mountainous regions, and offers a useful reference for high-speed grading of other flexible leafy agricultural products. • Designed a compact fresh tea leaf sorting device tailored for integrated harvesting and sorting equipment in hilly tea plantations. • Developed a lightweight fresh tea leaf sorting network based on regional segmentation combined with an ultra-simplified Transformer architecture. • Proposed a dynamic ROI extraction method based on region tracking and growth, significantly improving recognition accuracy for leaves crossing frame boundaries. • Achieved simultaneous sorting into three grades in a single detection, with real-time inference at 50–70 frames per second on a CPU-only platform and a single-channel sorting capacity of up to 26.7 kg/h.