Camellia Oleifera Fruit Detection Based On Computer Vision And Deep Learning Technology

Camellia oleifera has been planted in China for more than 2000 years.It is known as"Oriental Olive Oil"and has high nutritional and economic value.In the production of Camellia oleifera fruit,the labor cost accounts for 20～30%of the total harvesting cost.As China is undergoing industrial upgrading,a large number of people are pouring into the secondary and tertiary industries.At present,the number of people engaged in agriculture is getting smaller and smaller,and the labor shortage problem is becoming increasingly serious.However,the automatic harvesting of Camellia oleifera fruits with mechanized equipment can greatly reduce the labor cost,which is expected to alleviate the problem of labor shortage.The accurate positioning of Camellia oleifera fruits is the key to the correct work of Camellia oleifera fruit picking equipment.Based on the deep learning technology,this study developed a target detection system of Camellia oleifera fruit in unstructured scenes,which mainly includes an image enhancement subsystem and a target detection subsystem of Camellia oleifera fruit.The main research contents and conclusions are as follows:（1）To solve the problem that the existing low illumination enhancement algorithms were difficult to enhance the images of Camellia oleifera fruit in natural scenes,a low illumination enhancement algorithm for Camellia oleifera fruit（A-UNet）was proposed.A-UNet was a low-light image enhancement algorithm based on U-Net Convolutional Neural Network（CNN）and integrating channel attention mechanism.The experimental results showed that the peak signal-to-noise ratio（PSNR）,structural similarity（SSIM）,relative brightness order（LOE）and feature similarity（FSIM）of the image enhanced by A-UNet were 31.76,0.98,37.49 and 0.98 respectively.Using the dataset of Camellia oleifera fruit enhanced by A-UNet to train YOLOv4 model,the Precision（P）of the test set was93.85%,the Recall rate（R）was 89.63%,the average correct rate（m AP）of the whole class was 93.85%,and the F-Measure（F₁）value was 0.92,which were respectively higher than those of P,R,m AP and F₁of the control model（2）Aiming at the low recognition accuracy of common target detection algorithms when fruits overlap,a Soft-YOLO algorithm for identifying overlapping fruits of Camellia oleifera was proposed.Traditional target detection algorithms often used greedy NMS algorithm to remove the adjacent prediction frames in overlapping areas,which directly removed the frames with low scores in overlapping areas,which would easily lead to false detection and missed detection of overlapping fruits.When Soft-YOLO screened overlapping frames,it modified the hard zeroing strategy of NMS to set attenuation function,thus allowing the network to accurately identify the corresponding fruits in overlapping areas and increasing the detection rate of overlapping fruit targets.The experimental results showed that this algorithm could recognize the images of Camellia oleifera fruits with high accuracy in different overlapping degrees,and the m AP of this algorithm is 94.74%,which was superior to 93.43%of the original YOLOv4 network.It could meet the requirements of Camellia oleifera fruit harvesting robot for fruit positioning accuracy.（3）Aiming at the low recognition accuracy of Camellia oleifera fruit image in night environment,a night Camellia oleifera fruit recognition algorithm YOLON was proposed.On the basis of the target detection network YOLOv3,YOLON introduces the Illumination Adaptive Adjustment Module（LA）,which could adaptively adjust the illumination of the camellia oleifera fruit image at night,making the outline and details in the feature map clearer.In addition,this study put forward the Night Prior Knowledge Module（NPK）,which models the errors of the network,and finely adjusts the prediction results of the network feature map in the form of auxiliary factors,so as to improve the recognition accuracy of the network.The experimental results showed that the test m AP of Camellia oleifera fruit images in the YOLON model proposed in this study was 94.37%at night,which was better than the original YOLOv3,and could meet the positioning accuracy requirements of Camellia oleifera fruit harvesting robot.（4）Using PyQt5 to design the software system of Camellia oleifera fruit target detection.Combined with the proposed target detection method of Camellia oleifera fruit,a comprehensive image processing system of Camellia oleifera fruit in unstructured environment was designed.The illumination of Camellia oleifera fruit images collected in natural scenes was adjusted,and target detection was carried out for Camellia oleifera fruit images in different scenes.The experimental results showed that the system could enhance the illumination and detect the target of Camellia oleifera fruit images accurately and in real time.