The Research On Image Recognition Technology Of Camellia In The Natural Light Based On Convolutional Neural Network

The workload and the cost required to harvest camellia fruit will be reduced to greatly improve the work efficiency by the application of the picking robot.How to efficiently identify and detect natural images collected by sensors during the picking process has become an important technical node that restricts the application of picking robots.Therefore,research on the identification technology of camellia in natural light environment possesses an important practical significance and broad application prospects for the camellia industry.The paper focuses on the following aspects of research work:Aiming at the characteristics of camellia fruit images under natural lighting environment,analyzing the shortcomings of the existing methods of feature extraction and recognition classification,a classification network of camellia fruit recognition based on improved convolutional auto-encoder is proposed.On the basis of the convolutional auto-encoder,the feature extraction module is modified.The parallel decomposed convolution kernels are used to learning different types of features.The parallel decomposed convolution kernels are also used to improving the learning ability,recognition performance,speed performance and stability of the network.It provides the basis for the identification task of camellia fruits on the actual image.Based on the image recognition task flow of the picking robot,a method of detecting and segmenting the camellia based on the improved mask rcnn network is proposed.In view of the fact that the teafruit recognition task has low network depth requirements and high speed performance requirements,the improved convolutional auto-encoder network is used as a feature extraction network for improving the mask rcnn network,to improve the speed performance of the network to make it real-time.The pyramid feature network is used to fuse the context features to learn the feature images at different scales.On this basis,the regional recommendation network is used to generate candidate windows,which are classified and segmented to realize the integrated processing of detection,recognition,classification and segmentation of camellia fruit on the actual image.Combining the growth characteristics of Camellia oleifera,the diversity of illumination environment and the characteristics of the corresponding two networks,the image of Camellia oleifera was further processed to construct appropriate data sets.Using the above data set and comparison algorithm,combined with the image of camellia collected under natural lighting conditions,the simulation experiments of two networks were carried out.The experimental results show that the improved convolutional auto-encoder network has good recognition and classification performance and speed performance.The network structure also possesses the good stability.The improved mask rcnn network inherits the high performance of mask rcnn network in recognition,detection and segmentation tasks.And have further improvement in speed performance,with a certain real-time.At the same time,the detection and segmentation network was tested in the field.The experiment proved the feasibility of improving the mask rcnn network under actual conditions,and verified the practicability and effectiveness of the network.