Research On Clustered Pod Pepper Target Recognition And Location Algorithm Based On Deep Learning

There are problems of leakage and high damage of pepper in mechanical harvesting operation of pepper.The main reason is that the driver adjusts the lifting of cutting platform according to experience,and the cutting platform cannot adapt to the growing height of pepper cluster.In order to quickly and accurately identify pepper and locate its height off the ground,guide the cutting platform to rise and fall in real time,improve the net extraction rate and reduce the damage rate,this paper constructed an image recognition and spatial positioning scheme for clustered pod pepper fruit based on deep learning and deep vision.The pepper fruit height output by this scheme is the pre-input of the intelligent transformation of the harvester cutting platform.The main research contents of this paper are as follows:（1）Construction of RGB-D image dataset of clustered pod pepper.Real Sense D435i binocular camera equipped with depth camera,after dynamic calibration,is used to collect 328 valid RGB and depth images respectively through field experiments.Based on the idea of"cluster"annotation,RGB and depth images of pod pepper and fruit are jointly annotated and amplified to obtain a total of 3280 images in the dataset.（2）A deep learning base network suitable for the detection task of clustered pod pepper is preferably selected through comparison experiments of target detection algorithms based on RGB images.To improve the recognition accuracy of clustered pod pepper in unstructured growth environment,comparative experiments are conducted on Faster R-CNN,SSD,and YOLO series of mainstream deep learning algorithms.In those RGB-image based experiments for pod pepper recognition,YOLOv5 network gains the highest detection accuracy of 83.6%,with F₁score of 0.78.Its all indexes are higher than other models,which manifests that this model is more suitable for the detection task of tufted pepper fruit of pod pepper.（3）Based on the YOLOv5 preference network,the research of RGB-D image fusion-based target detection algorithm for clustered pod pepper is carried out.In order to make full use of Depth image information and further improve the pepper recognition accuracy,the RGB and Depth image data fusion processing method based on Efficient Channel Attention（ECA）channel attention mechanism and Swin-Transformer backbone feature extraction algorithm is proposed.The results show that:average recognition accuracy of the improved model reached 84.3%.Compared with the original YOLOv5 network trained only with color images,five evaluation indexes of m AP_small,m AR_small,m AP₅₀,m AP_0.50⁰.95 and F₁ value are increased by 2.2%,8.6%,0.7%,6.8%and 2% respectively.