Field Weed Recognition Based On Light Sum-Product Networks

The unprecedented development of artificial intelligence has brought dramatic changes to all walks of life,and the driving force behind it is inseparable from deep learning.In the agricultural field,weeds are one of the main reasons for restricting food production.Modern agriculture is gradually becoming intelligent,such as the use of plant protection drones and intelligent weeding equipment.The key is how to effectively identify crops and different weed communities,and then efficient use of herbicides or weeding equipment.At present,deep learning architectures represented by convolutional neural networks have achieved remarkable results in image recognition,but they often require a large amount of sample data and high-performance hardware equipment to support the training and operation of complex network structures.The acquisition of sample data,labeling cost,network size,and operating efficiency in the application are all factors to be considered.Therefore,a small sample-based,lightweight network architecture has become one of the current research hotspots.Sum-product networks are a new model of probabilistic deep network,different from Bayesian networks and Markov networks,which have NP difficult to calculate probability graph model,Sum-product networks can perform accurate reasoning in linear time of network size,and have strong theoretical support.Compared with neural networks,its internal structure has clear probability semantics,and the learning process is simple and fast.In view of the above problems,this paper proposes a Light Sum-product networks algorithm that generates a network structure from the bottom up and updates the network parameters from the top down.The optimal logarithmic likelihood was obtained in the comparison of multiple benchmark data sets.Finally,we propose a Light Sum-product networks weed recognition model which combine with K-means algorithm.Using the soybean weed small sample data set obtained by the Unmanned Aerial Vehicle as a training set,the network parameters are only 36% of the sum product network,and the average recognition accuracy is 96.4%.Without losing accuracy,fewer samples are needed and training time is faster.The main work of this paper is summarized as follows:1? The research background and significance of light Sum-product networks are described.In the field of agriculture,the main methods of field weeding in recent years,the current research status and their problems are introduced.2? Introduced the knowledge of sum product network theory and related definitions and theorems.3? The light Sum-product networks learning methods based on mini-batch learning are detailed.Structural learning is different from the traditional method of recursively dividing instances and variables.Its network structure is dynamically generated according to the correlation between variables,and the number of each layer and node is constrained so that it has a lower tree width and Deep layers.In parameter learning,using Bayesian moment matching to update the weights from the top to the bottom and further optimize the model structure and parameters in a certain way.4? In order to verify the validity of the light Sum-product networks in the probabilistic reasoning task,a comparative experiment was performed on a large number of benchmark data sets,and good results were achieved in the average log-likelihood index.5? Weed recognition models with K-means as the underlying feature extractor and light Sum product networks as the high-level feature extractor and classifier are proposed.Comparing the traditional methods LearnSPN and AlexNet in a small sample data set of soybean fields,the model has higher accuracy and recall,and its performance is further analyzed and compared to verify the validity of the model.