Research On Compost Maturity Identification Based On Deep Learnin

Traditional compost maturity evaluation uses physical,chemical,and biological indicators for assessment,which involves complex steps,long measurement times,and high workload,lacking operability.This thesis proposes to study the intelligent recognition of compost images from the perspective of deep learning,and fuse multiple pieces of information such as images,temperature,humidity,and p H values to improve the accuracy of compost maturity recognition.Based on theoretical research,this thesis develops an intelligent recognition system for compost maturity.The main research content is as follows:(1)Collect and gather compost images.Collect compost images and temperature,humidity,and p H value data throughout the entire compost cycle,and collect compost images from different raw materials and origins,including images of compost made from animal manure,vegetable scraps,and straw,to expand the sample size of compost images.A total of nearly40,000 image samples were used,which were abundant and diverse.(2)Design a maturity recognition algorithm based on the CoAtNet model.This thesis adopts an improved CoAtNet model to design a maturity recognition algorithm,which extracts multilevel and multi-scale compost image features,constantly learns and adjusts network model parameters,perceives changes in color,texture,particle size,etc.in compost images at different periods,and modifies convolution layers and attention mechanisms to enhance the classification performance of the network model on the compost image dataset.The accuracy of compost maturity recognition reaches more than 99%.(3)Design of a multimodal fusion algorithm for maturity recognition.In this thesis,a multimodal data recognition algorithm model is constructed by fusing composting images with temperature,humidity,and p H values.The lightweight Mobile Net V3 network model was selected to extract image features,and the data features were extracted by stacking fully connected layers and non-linear activation functions.Finally,the image features and data features were combined by adding their dimensions and then classified.Performance is improved by increasing the channels.The use of multimodal data enriches the sources of data for the model,providing more relevant information on compost maturity for model training.(4)Develop an intelligent recognition system for compost maturity.Use STM32 as the main control chip to develop an automatic collection and uploading system for compost images and sensor data,and use Py Qt as the image interface development tool to design upper computer image recognition and monitoring software.Different maturity recognition algorithm models are loaded according to specific task requirements and actual situations to enable users to intuitively see the detection status and results of the system,and feasibility testing of the system is conducted in actual scenarios.Transfer learning can be used to train the network to adapt to compost maturity judgment with different raw materials.The designed portable collection device realizes rapid determination and recognition of compost maturity under multiple indicators,providing guidance for industrial compost maturity recognition.