Research On Apple Target Recognition Algorithm Based On Lightweight Deep Learning Network

China is one of the world’s leading apple-growing regions,but currently the domestic apple industry relies on manual picking,which requires a lot of labor and is slow,while workers have certain risks in climbing and picking.A fast and accurate automatic apple picking system can accelerate the intelligent development of apple industry,and the computer vision-based apple detection is an important part of the automatic picking system,which has a very far-reaching significance for the development of automatic apple harvesting equipment.In this study,a lightweight deep learning based apple fruit detection model in natural environment is proposed to achieve fast and accurate fruit identification,and the main research contents and conclusions are as follows:（1）Construction of apple dataset under natural environment.First,fruit images are collected in orchards on sunny days,cloudy days,different lighting conditions,and different shading methods.In order to enrich the image dataset and avoid overfitting,offline enhancement and online enhancement of image data are performed by data augmentation series to improve the generalization of the model.Finally,the apple fruit image dataset with 640×800 pixels size is obtained by data annotation.（2）Improved FCOS network-based apple fruit target detection in natural environment.In order to accurately locate apple fruit targets in natural environments,an improved full convolutional single-stage anchorless frame network for apple target detection is proposed.The network uses Darknet19,which has a smaller model size,as the backbone network based on the traditional FCOS network,and introduces the center-ness branch to the regression branch.A loss function that incorporates joint cross-ratio and focal loss is also proposed to improve detection performance while reducing the error caused by the imbalance between positive and negative sample ratios.The study conducts detection experiments on a computer workstation and compares and analyzes the detection results with those of the conventional FCOS network.The average detection accuracy based on the improved FCOS network is 96.0%,which is 3.1%higher than that of the traditional FCOS network,and the mean detection accuracy is 96.0%,which is 2.8%higher than that of the traditional FCOS network.The experimental results show that the improved FCOS network has better detection accuracy and stronger robustness than the traditional FCOS network.（3）Slim-FCOS network-based apple target detection in natural environment.A SlimFCOS lightweight apple target detection model is proposed for the FCOS network model that is computationally intensive and difficult to deploy to resource-limited embedded devices.Firstly,the simple structured Darknet19 is used as the backbone network to extract fruit features,the number of channels of the detection head is reduced to 128 dimensions,and the number of channels of the detection head is further reduced by replacing the normal convolution of the detection head with the depth separable convolution,and the path aggregation network is used instead of the feature pyramid network.Meanwhile,to further simplify the network model and ensure the detection efficiency,the FCOS is pruned using the BN-based channel pruning algorithm to reduce the amount of operations,and finally the pruned model is fine-tuned using the CIoU border regression loss to achieve fast and accurate detection of apple fruits.The experimental results show that the Slim-FCOS detection model achieves effective detection of fruits with different light and different densities with an average accuracy of 95.4%.The model size decreased by 70.2 MB compared to the FCOS network,and the model parameters were reduced by 62.3%.The average detection time of a 640×800 image at the workstation was 27.5 ms,and the method reduced 6.8 ms,2.0 ms,24 ms,and 0.8 ms slower than YOLOv4,YOLOv5s,and improved FCOS,respectively.the average detection accuracy improved by 1.6,0.4,and 0.6 percentage points,respectively.slim-FCOS Slim-FCOS effectively simplifies the model with guaranteed detection accuracy,and provides technical support for model migration to mobile with limited computing power.（4）Model deployment.Apple detection models are deployed using an edge device,Jetson Xavier NX,which is a device with very good computational performance and very low power consumption.Firstly,the system is burned on the embedded device,the data transfer,model weights and code migration are realized by winSCP software,and the deep learning environment is deployed on the device,and finally the model validation is performed on the embedded device.The experimental results show that the average detection accuracy of the model of Slim-FCOS network is 95.4%,which is close to the improved FCOS and higher than YOLOv4,YOLOv5s and YOLOX.The average time to detect a single image is 140.3 ms,which is 170.2 ms,30.2 ms and 10 ms faster than the detection times of the improved FCOS,YOLOv4 and YOLOv5s networks,respectively,and close to the detection time of YOLOX,and the Slim-FCOS network model achieves fast fruit detection on edge devices.