Research On The Ear Loss Detection System Of Corn Harvester Header Based On YOLOX

At present,the mechanized harvesting rate of maize in China is about 70%,which is lower than the mechanized harvesting level of wheat and rice.Harvest mechanization is an important means to promote the cost-saving and efficiency-increasing of maize industry and improve the comprehensive economic benefits and international competitiveness of maize production in China.The working performance of corn harvester directly affects the working efficiency and harvesting quality.The establishment of intelligent detection system for corn harvester is an important means to improve the working performance of harvester towards intelligent harvesting.The real-time monitoring of ear loss in corn header is an important part of intelligent detection fault detection.At present,the loss of corn ears in the header is mainly based on manual observation.The efficiency of manual observation is low,the amount of labor is large.and it cannot be detected in real time.Therefore,this paper studies the corn ear loss detection system based on YOLOX,which is used to realize the real-time detection of corn ear loss.The main research contents are as follows.(1)The maize ear data set was constructed and the image acquisition was carried out in the experimental field of Shandong Agricultural University.The maize ears lost in the field at different angles were taken,including the peeled ears and the unpeeled ears.The collected images were preprocessed,image amplified,and image annotated,and a self-built corn ear data set was established.(2)The mainstream YOLO series target detection networks YOLOV3,YOLOV4,YOLOV5 and YOLOX are analyzed,and the self-built data sets are transferred and learned.Through the analysis of the results of transfer learning,the algorithm YOLOX with the best detection effect and detection speed is selected for optimization research.The selected YOLOX algorithm is optimized.Firstly,the GAM attention mechanism is selected to optimize the model to improve the feature extraction ability of the original model,and compared with the SE and CBAM attention modules.The improved YOLOX loss function replaces the IOU loss function with the DIOU loss function to improve the convergence speed of model training;the improvement of the activation function improves the reasoning ability of the network model,and proposes RBF-bottleneck to improve the NECK structure of YOLOX.The FPN structure is changed to the RBF-PANT structure,which increases the receptive field and improves the feature fusion ability of the YOLOX network model.Finally,the ablation test of the model is carried out,which proves that the improved model can combine the advantages of each improved module,improve the detection speed,recognition accuracy and reduce the training time to accelerate the convergence of the model.(3)The network model is pruned and compressed to determine the sparse rate and the optimal pruning rate is determined by pruning under the sparse rate.Through experiments,the selected sparse rate can ensure higher accuracy while reducing the model.(4)The GUI interface design of the corn ear loss detection system is based on Python and Py QT5,which can realize the interaction of the human-machine interface.It is verified that the interface design is simple and easy to operate to meet the real-time detection requirements,and then the designed detection system is deployed to Jeston Nano.(5)The corn ear detection bench test was carried out.The single factor test was carried out to determine the value range of each factor detection effect for orthogonal test.Secondly,the orthogonal test scheme is determined according to the value range of each test factor.Finally,the range analysis is carried out according to the test data to determine the influence degree of the three factors of conveyor belt speed,camera height and camera angle on the recognition effect.The conveyor belt speed has the greatest influence on the detection effect,and the camera angle has the least influence.The optimal combination of detection effect is determined as when the conveyor belt speed is 6m/s,the camera height is 0.6m,and the camera angle is 0 °.