Research On Persimmon Inspection Recognition And Fruit Nondestructive Testing Based On Deep Learning

The accurate and efficient identification of target fruits plays an important role in agricultural production links such as intelligent picking,growth monitoring,and fruit detection.However,in the natural orchard environment,different lighting,weather conditions,fruit overlapping,blocked by branches and other complex factors bring certain difficulties to the identification and detection of fruits and the determination and grading of fruit quality.In this thesis,persimmons are taken as the research object,and the inspection and identification of early green persimmons is carried out,and the fruit detection of persimmons in the color transition stage is realized from the two aspects of ripeness grading and fruit size measurement,so as to improve the non-destructive testing of persimmon identification accuracy and fruit quality in natural environment.The specific contents are as follows:(1)After data preprocessing and labeling,persimmon images were collected in the field and divided into two data sets,which were used for multi-cluster green fruit study and fruit maturity grading study respectively.And provide data basis for the subsequent algorithm research.(2)Multi-cluster green persimmon identification method based on DEM-Faster RCNN.Det Net is used as the backbone feature extraction network to add a weighted ECA channel attention mechanism to the three effective feature layers in the Det Net network.The lowlevel features are maximally pooled so that they are the same dimension and magnitude as the high-level features.The processed feature layer uses a serial hopping layer connection structure for multi-scale feature fusion.The K-means clustering algorithm is used to cluster the bounding box and the anchor box,so that the anchor box tends to the real bounding box.The m AP of the DEM-Faster RCNN model reaches 98.4%,which is 11.8% higher than that of the traditional Faster RCNN model,and the average detection time of a single image is increased by 0.54 s,which is a significant improvement in accuracy and speed.(3)Mask RCNN persimmon ripeness detection algorithm optimization based on encoder.The color component of the standardized maturity grading detection is compared by the chromatic difference method,and the multi-channel image is constructed by combining the effective color component map.The encoder is used to input this image into the Ro I Align structure to realize the combination of high and low resolution feature images.The total m AP value of the improved object detection algorithm reaches 92.71%,and the total missed detection rate and total false detection rate are reduced by 4% and 17% compared with the traditional algorithm.The improved Mask RCNN can achieve accurate judgment of persimmon maturity without being affected by external environmental influences,and realize the detection and identification of small targets through the combination of high and low resolution.(4)Mask RCNN instance segmentation algorithm optimization.After acquiring the object detection image in the fully connected layer after the Ro I feature,the cropping,morphological processing and pitting segmentation modules are added to the mask head.Focus on the target object,and then further divide to solve the adhesion problem caused by overlapping and blocking the fruits.The experimental results show that the m Io U and m AP value of the improved instance segmentation algorithm are 92.7% and the m AP value is94.25%.The overall relative error range of fruit diameter was between 0.97%～1.13%,which improved the segmentation accuracy of the target fruit and reduced the relative error value of subsequent fruit diameter measurement.