Research On Detection Algorithm Of Industrial Equipment Components Based On Small Target Detection

Industrial intelligence is growing with the rapid development of artificial intelligence and deep learning.Target detection technology is also gradually applied to fields such as autonomous driving,intelligent monitoring,intelligent transportation,and defect detection,which significantly improves production efficiency.The research of small target detection with small targets,weak features,and lack of sufficient appearance information is also developing in-depth,and there are also relevant application scenarios,such as aerial remote sensing image detection,battery defect detection,etc.For the industrial equipment component detection scenario,this thesis uses automated industrial equipment component detection to replace the traditional inefficient,costly and error-prone manual detection,which helps improve detection efficiency,reduce labor costs,and lower error rates.The main work of this thesis is as follows:1.Collect and build a dataset of industrial equipment components: Based on the collected and labeled data of signal equipment components in signal rooms of high-speed railway stations and electrical components of substation switchgear,an industrial equipment component dataset is constructed to support the training requirements of the model in this thesis.And to solve the problems of the small amount of existing data,unbalanced positive and negative samples of data,and small targets to be measured,this thesis uses various data enhancement strategies applicable to small targets,such as cropping and splicing,horizontal flipping,replication enhancement,modification of contrast and brightness,random noise and random erasure,to augment the data set for improving the robustness of the model.2.Propose ASFF-HRNet,which is an improved model base on the HRNet model for industrial equipment component detection: In this thesis,we propose the use of an adaptive spatial feature fusion method for the problem of inconsistency between features at different scales after multi-scale feature fusion,aiming at the model adaptively enhancing the weights of more important features to improve the accuracy rate of the model.Through the comparison experiments with the singlestage target detection model representative YOLOv5,the two-stage target detection representative Faster R-CNN,and the baseline model HRNet,which is often applied to small target detection,it is found that the ASFF-HRNet model achieves the best detection performance with 4.3%,2.27%,1.1% improvement in m AP values compared to the above three models,respectively which proved the effectiveness of the improved scheme.3.Propose industrial equipment component detection algorithm base on the point set representation: To solve the problem of missed detection that often occurs when small targets are aggregated,and to further fine-grained detection of target components,this thesis proposes an anchor-free target detection strategy using representative points based on the ASFF-HRNet model,which contains important local semantic information for better classification while the representative point set is more fine-grained to detect the target object location.Through experimental comparative analysis,the improved model improves in comprehensive performance over ASFF-HRNet,demonstrating that the performance of the improved model on the industrial equipment component detection problem in the industrial equipment component dataset exceeds the performance of the currently common general-purpose target detection models on this task.4.Designed a software prototype system for industrial equipment component detection: The system mainly includes a front-end,server-side,and edge-side,the front-end mainly provides visual display pages and interactive interfaces for inspectors,the server-side mainly provides several interfaces for the front-end,and controls the tasks of the edge-side and maintains historical data,the edge-side is responsible for responding to the tasks of the server-side and real-time detection of the collected images and saves the detection results into the database of the server-side.The ASFF-HRNet model,the representative point-based industrial equipment component detection algorithm,and the software prototype system for industrial equipment component detection proposed in this thesis achieve better detection results on the selfbuilt dataset,which is conducive to the intelligence of industrial equipment component detection.