Research On Strawberry Detection In Greenhouse Based On Improved YOLOv5 Algorithm

At present,China has become the world’s largest strawberry producer,and the planting area and production of strawberries have been increasing year by year,but the yield estimation and fruit harvesting of domestic strawberries still mainly rely on manual labor,which is time-consuming and labor-intensive.With the development of agricultural modernization,strawberry picking robots and intelligent information monitoring technology have gradually become a research hotspot,among which the accurate detection of ripe and unripe strawberries is one of the key aspects to achieve strawberry yield estimation and automatic picking.However,strawberries grown in natural environment are denser and there are many small target fruits,and feature information is difficult to be fully extracted,which causes difficulties for accurate identification of strawberry fruits,and the existing convolutional neural network target detection models are usually designed to be more massive and have poor real-time detection,which is difficult to meet the actual orchard application requirements.In response to the above issues,this thesis takes the strawberry images collected in greenhouse as the research object,carries out research from two aspects of improving the fruit recognition accuracy and lightweight model,and proposes a strawberry detection method based on improved YOLOv5 to achieve fast and accurate recognition of strawberries in greenhouse.The main research contents and conclusions are as follows:(1)From the perspective of improving strawberry detection accuracy,an improved Accurate-YOLOv5 model was proposed.The model highlighted important features of strawberry by adding Coordinate Attention(CA)mechanism to the backbone network to capture cross-channel information,orientation-awareness and position-sensitive information,and introduced Adaptive Spatial Feature Fusion(ASFF)module after the neck network to solve the inconsistency problem between multi-scale features,suppress conflicting features at different scales and enrich semantic information in the shallow feature layer to improve the recognition of small targets,and SIo U Loss was used as the loss function of the bounding box regression to stengthen the target positioning ability and improve the accuracy of the bounding box regression.The research results showed that the proposed Accurate-YOLOv5 model achieved a mean average precision of 95.19% on the test set,a 4.39%improvement compared to the initial YOLOv5 s model.The improved model also outperformed mainstream high-accuracy algorithms such as Faster-RCNN,Center Net,YOLOv4,and YOLOX in terms of detection performance.(2)A Light-YOLOv5 model based on lightweight improvement was proposed for the problem that the parameter volume of the improved network become larger and the computational cost become higher.This model designed a lightweight Ghost Net network structure for feature extraction in the backbone.GSConv was used to build the Slim-Neck network structure,replacing the standard convolution responsible for downsampling in the ASFF module.This design reduced the parameter and computational complexity of the feature fusion structure while maintaining the model’s recognition accuracy.The research results indicated that the proposed lightweight design effectively reduced the network parameters.The mean average precision of the improved Light-YOLOv5 model on the test set was 93.65%,the number of network parameters and floating point calculation amount were 6.06 M and10.53 G,and the average detection time of each image was only 7.21 ms.The model outperformed the initial YOLOv5 s model in all aspects.Compared to lightweight algorithms such as YOLOv4-Tiny,YOLOX-Tiny,YOLOv4-Mobile Netv2,and YOLOX-s,this model achieved a good balance between recognition accuracy,detection speed,and network parameters,making it more suitable for porting to low-performance embedded devices.(3)Application research of lightweight Light-YOLOv5 Model.Firstly,the burning of the system image,the initial setting and the configuration of the deep learning environment were completed on the embedded device of Jetson Nano.Then,the improved model weight was transplanted to the device and accelerated processing was carried out.Finally,the strawberry detection system in the shed was designed on this hardware platform.Py Qt5 and Qt Designer were used to make the user information page and detection homepage,and various functional options were provided on the detection homepage to realize the detection and counting of ripe and unripe strawberries.In summary,the Light-YOLOv5 algorithm proposed in this thesis can achieve fast and accurate identification of strawberries in greenhouse with less number of parameters and computational effort.The research results can provide accurate target information for the intelligent monitoring of strawberries,facilitating yield estimation,and offering important theoretical support for the development of vision systems in fruit-picking robots.