| With the rapid development of artificial intelligence and agricultural robot technology,traditional agriculture is gradually transforming to intelligent,and the intelligentization of agricultural production will become the main trend of modern agricultural development.As an important agricultural by-product and economic crop in my country,tea has not yet achieved intelligent production.The automatic picking of tea leaves is the core issue to realize the intelligent production of tea leaves.At present,ordinary bulk tea leaves can be picked by nonselective tea picking machines,but high-end famous tea leaves mainly rely on manual work.The main reason is that famous and high-quality tea leaves need to be selective.of picking.Due to the different growth postures of tea buds in natural scenes and the complex and changeable environment,there is still a lack of reliable recognition and picking point positioning methods to replace manual work.For this reason,based on deep learning and traditional image processing methods,this paper studies the recognition and picking point positioning methods of famous and high-quality tea buds in natural scenes.The main research contents and relevant conclusions are as follows:(1)Research on tea buds rotated object detection algorithm.First,analyze the picking requirements of famous and high-quality tea buds,determine the recognition of single bud tip and one bud and one leaf,and divide them into four detection classes according to the growth posture;Second,establish a tea buds rotated object detection dataset,and analyze the difficulty of identifying tea buds to enhance the dataset;then through the analysis of the algorithm principle,the Oriented R-CNN algorithm based on the two stages is selected for research and improvement;Aiming at the problem of low detection accuracy of the Oriented R-CNN algorithm,an improved T-Oriented R-CNN is proposed Algorithm,research and improvement from two aspects of feature extraction network and feature fusion network: Combining the characteristics of Res Ne Xt50 network,CSPNet network and SPP spatial pyramid pooling module,a feature extraction network with low complexity and high performance is proposed;based on CA attention mechanism proposes a high-performance feature fusion network.Finally,through experimental comparison,it is verified that the T-Oriented R-CNN algorithm greatly improves the detection accuracy when the detection speed is slightly reduced,and can realize high-precision recognition and rotated frame positioning of four detection classes of tea buds.(2)Research on the location method of tea bud picking point integrated with semantic segmentation.First,based on the rotated object detection dataset,four single bud semantic segmentation datasets with different detection classes are constructed;Then,in order to improve the segmentation speed of the semantic segmentation algorithm Deep Labv3+,a fast lightweight algorithm M3-Deep Labv3+ is proposed;further,the image surrounded by the rotated frame obtained in the detection stage is cropped to form a single bud image,and it is segmented by the M3-Deep Labv3+ algorithm to obtain a single bud mask image;Then further design the traditional image processing algorithm to process the mask image to realize the positioning of single bud picking point;and mapping the picking point back to the tea image;finally,the depth map is obtained through the 3D camera,and the positioning of the picking point of the tea buds is realized by combining the depth value.Finally,by comparing the M3-Deep Labv3+ algorithm with its basic algorithm experiment,it is verified that the M3-Deep Labv3+ algorithm greatly improves the segmentation speed while improving the segmentation accuracy slightly,and the model complexity is greatly reduced;By combining the algorithms used in this paper to form an algorithm for tea buds recognition and picking point location,and statistically analyzing the success rate of picking point location of the algorithm,the experiment shows that the success rate of picking point location for the four detection classes of tea buds is 89.9%,83.5%,88.9% and 87.3%;At the same time,it is tested that the algorithm takes an average time-consuming of 35.03 ms to identify a single tea bud and locate the picking point;it shows that the algorithm in this paper has high positioning speed while the positioning accuracy meets the requirements.(3)The design and implementation of tea buds recognition and picking point positioning system.First,analyze the application requirements of the vision system,and design the overall scheme of the system;second,select and build the hardware platform of the system;then for its software system,analyze the functional requirements of the software system,design and implement the software system,and further test the software system function;Finally,the experimental verification of the tea buds recognition and picking point positioning system shows the feasibility and advancement of the system.Based on the above,this paper studies the precise and fast recognition and picking point localization methods of tea buds in natural scenes.Aiming at the detection of tea buds,a highprecision rotated object detection algorithm suitable for tea buds is improved;On this basis,an efficient lightweight semantic segmentation algorithm is proposed,combined with the designed traditional image processing algorithm and the depth map obtained by the 3D camera,the rapid positioning of the picking point of tea buds is realized;Further combining the above algorithms to form a tea buds recognition and picking point localization algorithm,achieving precise and fast recognition and picking point localization of tea buds;Finally,a tea buds recognition and picking point positioning system was designed and developed;To lay a theoretical foundation and practical experience for the intelligent picking of tea buds. |
