Research And Development Of Log Tally Platform Based On Artificial Intelligence

The combination of artificial intelligence and actual production applications will be a new trend of innovation and development in the future;log tally is a very important business in the port.The current tally mode needs to be carried out on site by manual points,and the workload is very large;Due to the harsh on-site environment,there are hidden safety hazards,and a large number of personnel are required.In addition,the specifications of the logs are different,and the accuracy of the tally data and the controversial data cannot be guaranteed.In response to the above-mentioned series of problems,we have developed a set of log intelligent tally platforms combined with artificial intelligence algorithms.Through the platform,you can perform real-time effective tracking monitoring and logical status analysis of trailers and forklifts on site,as well as statistical analysis of the entire log loading and unloading process,so as to obtain the number of logs for each full fork,the number of logs for each trailer,and the number of logs for the entire ship;at the same time,the platform can display the analysis results in real time.The overall platform can be divided into GPU server and web server.GPU server is used to process video streams and images,detect and identify forklifts,trailers and logs,analyze the status and count the number of logs;web server is used for display,query,storage analysis,as a result,a small amount of manual intervention can be performed if necessary.The log intelligent tally platform integrates cutting-edge artificial intelligence algorithms,making the detection of logs easy to implement,but it also faces challenges;due to the different specifications of logs and outdoor operations,the robustness requirements of the model are very high,not only to adapt to changes in light,but also to detect dense and thin logs.Based on such characteristics,the deep convolution model was redesigned;in order to adapt to the characteristics of dense and small targets,it is necessary to optimize the receptive field of the model and make full use of shallow texture features;the main innovations and work are as follows:1.Propose a global context module and Four-branch feature fusion model with residual branching and deconvolution.Modeling through global context information,integrating feature space dependencies,and calibrating through channel dependencies,to achieve integration of context information and optimize log target edge information.One of the residual branches is used to transfer the features of the backbone feature layer.There are two branches that use convolution branches in different directions in space to further extract features.The deep features are upsampled by deconvolution as a residual branch.Finally,the features of the four branches are fused.2.Propose the pooling layer feature pyramid and the adjacent feature layer fusion module.The pooling feature pyramid is to achieve bottom-up weight sharing,plus topdown feature fusion,so that the entire network can make full use of the feature information of each layer,thereby accelerating the convergence trend of training losses,and in Small data concentrates on certain advantages.The adjacent feature layer fusion is further optimized.In order to reduce the feature overlap caused by different scale feature information and buffer the gradient flow during training,only the adjacent feature layers used for prediction are fused;the above optimization can further improve the training efficiency and test effect.3.Propose the use of fusion of different receptive field features for intensive log detection.In order to solve the problem that it is difficult to accurately locate due to the dense arrangement of small logs,according to the characteristics of hollow convolution,increasing the local detail information by increasing the acceptance field and resolution is more beneficial to the detection of the target without adding any overhead.In tracking analysis,the angle of the target changes in real time,resulting in serious occlusion problems between different angles and dense targets,especially small targets.Increasing the local receptive field can improve the detection efficiency of dense targets.4.Demand analysis and scheme design of the log intelligent tally platform.In the early stage of development,detailed requirements analysis and scheme design were carried out;key analysis was made on the operation process,interface interaction,function design,and web design,and development was carried out based on this standard.The purpose of the platform design is to integrate the algorithm with the live video stream,and display and save the identification and analysis results of the algorithm,while providing an interface to the original port system.We collected data sets of log sections under different specifications,different operating conditions,and different angles on site.Through this data set,through a large number of experimental comparisons,to show the advantages of the new model in practical operations,especially in the detection and identification of dense small-sized logs.