Research And Implementation Of Fine-grained Material Recognition Algorithm Based On Deep Learning

In recent years,the construction machinery industry has been facing new trends of greenization,intelligence,and high-quality development.Research has found that in the operational cycle of construction machinery such as loaders and excavators,the material digging process consumes the largest proportion of energy.During the construction of intelligent systems,different types of materials affect the power matching,operation trajectory,and work mode of construction machinery.Therefore,achieving high-precision identification of different material types is crucial for efficient operations.However,the small differences in characteristics between different materials and the harsh working conditions significantly impact the identification results.To achieve high-precision material identification and mitigate the impact of low-quality factors in adverse environments,this paper comprehensively applies deep learning algorithms to focus on fine-grained features,accomplishing high-accuracy image classification and object detection,and implementing relevant models.The main research work includes:(1)Research on high-performance fine-grained classification networks: Firstly,focusing on fine-grained feature aggregation,a novel Bilinear CBAM(BCBAM)with dual-line attention mechanism is constructed and designed.Secondly,the Conv Ne Xt-Tiny is used as the feature extraction network,and attention embedding schemes based on single-cycle and multi-cycle structures are designed.Finally,a multi-scale,multi-angle,and all-round attention framework is proposed and applied to the backbone network to construct BCBAM＿Conv Ne Xt.The feasibility and superiority of the network model in fine-grained classification are verified through comparison and validation on internationally recognized fine-grained classification datasets,providing a foundational network for fine-grained image classification of materials.(2)Conv Ne Xt-based fine-grained material image classification algorithm: Firstly,the potential low-quality factors in adverse environments are analyzed,and high-and low-quality material datasets are established to train and validate the network model’s classification performance.The influence of low-quality factors on the model’s performance is analyzed.Secondly,St＿BCBAM is introduced to enhance the network’s filtering ability for low-quality factors by utilizing the high-dimensional threshold groups adaptively provided by BCBAM combined with soft thresholds.Experimental results show that the improved model achieves an average accuracy of 97.9% and 97.7% for low-quality materials in different subcategories within the same main category and different subcategories of different main categories,respectively.(3)Optimization algorithm for fine-grained object detection based on YOLOv5:Firstly,high-and low-quality datasets suitable for object detection are established for fine-grained materials.Secondly,BCBAM and St＿BCBAM are improved to adapt to the data processing method of YOLOv5.An attention optimization structure based on YOLOv5’s feature extraction,feature fusion,and feature integration processes is proposed and studied through comparative analysis of the three optimization structures on high-and low-quality datasets.Experimental results show that the improved model achieves detection accuracies of 93.2% and 90.2% for high-and low-quality materials,respectively.(4)Implementation of image classification and object detection models: Dynamic and static models are constructed,improved,and designed,and anti-misjudgment mechanisms are created for different algorithms to ensure the stability of result uploads.This research delves into the fields of image classification and object detection,aiming to improve the detection performance of models.It innovatively optimizes algorithms,rectifies network structures,studies the best solutions,and combines the analysis of fine-grained material classification and detection effects.The proposed improvement methods and optimization algorithms achieve optimal recognition effects,providing technical support for the design of strong feature extraction networks in this field,as well as program design and application development in practical projects.