Reserach On Apple Young Fruit Detection And Location In Complex Environment Based On Deep Learning

The realization of orchard automation is the key to solve the problem of labor shortage and high planting cost.The detection link of the automatic process in the stage of the young apple was studied,and the detection and location of young fruit in complex environment was solved according to the methods of target detection and binocular recognition.The main work and contributions algorithm were summarized as follows:(1)Aiming at the problems of overlapping fruits and complex background of young apple fruits,a method of coarse and fine grain labels was proposed to detect young fruit targets according to the principle of target granularity.The coarse grain label of young fruits was set as the thinning area,which solved the problem of overlapping fruits.The fine grained young fruit labels was set to stop pruning and predict apple yield.Aiming at the problem of small target,a method of far and near detection was proposed according to the position relationship between the camera and young fruit.The distant-view detector that had wider receptive field collected young fruit information,and provided target positioning information for the lightweight close-view detector through the near-far conversion model.Then,the data set was established through data collection,data cleaning,data calibration and data enhancement.And the self-learning network framework was proposed to realize self-adaptability of the data set.Finally,the experimental analysis shows that YOLO(You Only Look Once)model was the most suitable for young fruit target detection.Based on this model,the superiority of data set separation method was verified,and the two image processing methods improved the precision of close range young fruit detection,the m AP value was increased by 3.78% and 3.76% respectively.(2)Aiming at the problem of low precision of distant-view detection,the distant-view detection methods(YOLOV5l-M)were improved to enhance the feature extraction ability of small target by proposing CBM＿G module and SPPE-SE module with SE channel attention mechanism according to the differences of young fruit feature channels,and improved the feature extraction,module.Aiming at the real-time performance of close-view detection,a lightweight close-range detection model(YOLOV5s-G)was improved based on CBL＿G43 module,SPPE-SE module and deep separable convolutional network.Experimental results showed that the accuracy of young fruit detection was improved by 5.6% in YOLOV5l-M network.The YOLOV5s-G detection speed was increased to 73 frames per second,the model size was reduced by 10%,and the detection accuracy was increased by 1.44%.(3)Aiming at the problem that the detection environment of young fruits was not conducive to depth localization,a method was proposed to constrain the location of SAD young fruit.Based on discrete cosine transform,a perceptual hash function method was proposed to constrain the stereo matching process,and the search range of SAD algorithm was restricted to the vicinity of the target center point,so as to accelerate the detection speed and improve the accuracy of young fruit detection.The experimental results showed that the accuracy of the constrained SAD localization method could meet the requirements of the detection of young fruits,and realized the accurate identification and localization of young fruits.