Impurity/Breakage Assessment of Vehicle-Mounted Dynamic Rice Grain Flow on Combine Harvester Based on Improved Deeplabv3+ and YOLOv4

It is difficult to extract small and dense objects with random state, such as grain impurity, in image of vehicle-mounted dynamic rice flow on combine harvester. Therefore, this paper improves Deeplabv3+ by constructing MobileNetv2 coding layer adding ECA(Efficient Channel Attention) Encoder Decoder improve extraction accuracy high-dimensional features images a large number state. In addition, the YOLOv4 improved using Mixup preprocessing, Mish Neck Head, ECA Prediction BackBone training precision reducing effect gradient disappearance. And impurity/breakage rates are assessed based relationship model between pixel area quality, YOLOv4. The proposed method was verified experiments acquired intelligent Compared existing Deeplabv3+, YOLOv4, U-NET, BP, increased more than 4.01%. average relative error time assessment were 7.69% 1.56s. can accurately rapidly assess for harvester, further realize closed-loop control harvesting operation.