Field Strawberry Ripeness Recognition Method Based On Deep Learning

Strawberry is a common fruit in Chinese fruit market.It is loved by the market because of its rich nutrition and low content of fat.At present,the planting area of strawberry in China ranks No.1 in the world.The increase of strawberry plant yield will lead to high fruit growth density,easy to cause strawberry fruit overlap,and large and dense plant leaves will block the fruit,which will bring challenges to strawberry automatic picking and affect the accuracy of strawberry target recognition by computer vision technology.In addition,strawberry maturity classification is still in its infancy,which is limited to manual classification and traditional classification based on color features.It is easy to be disturbed by environmental factors while extracting features.Therefore,the main research contents of this paper are as follows based on the above problems:（1）A set of field strawberry data sets were established,and then the images were sharpened to improve the edge information of strawberry target.The data set was expanded by data enhancement to expand the richness of the data set.（2）The Faster R-CNN model was improved by adding feature pyramid network to enhance the recognition accuracy of the model for small target objects.The soft non-maximum suppression algorithm was used to enhance the recognition accuracy of overlapping targets and reduce missed detection.The parameters of regularization modes,anchor points and learning rates of the model were adjusted to make the model more suitable for the target recognition task of field strawberry.The experimental results showed that the improved model had a better performance than the Faster R-CNN and YOLO-v4 models while recognizing strawberry target in field.The mean average precision was improved by 7.39%and14.49%,and the F₁ score was increased by 16.57%and 15.17%,respectively.The different light and target number experimental results showed that too dark or too bright data would reduce the accuracy of target detection,and the less the number of targets,the higher the accuracy of detection results.（3）A lightweight strawberry ripeness grading network（SRGNet）was designed to replace manual grading and reduce the subjective error.Combined with the results of target detection task,a strawberry maturity classification data set was established according to the obtained prediction frame location to train the SRGNet.The experimental results showed that the SRGNet had higher accuracy than the traditional feature extraction network,up to 98.71%.The SRGNet run 7 times faster than other networks in terms of the test frame rate.