Research And Development Of Intelligent Detection System For River Floating Objects

As a common problem in river pollution control,more and more experts and scholars carry out intelligent detection of river floating objects.Based on the river video captured by UAV(unmanned earial vehicle),a method of river floating object recognition based on SSD(Single Shot Multibox Detector)depth network is proposed.SSD uses the pre-training network model of COCO dataset for migration training.Based on the floating object detection algorithm,an intelligent floating object detection system based on Android is designed.In order to realize the detection system,the main work of this paper is as follows:(1)Dataset production of river floats.First of all,the traditional image processing technology is used to design an algorithm for river videos with severe pollution from floating objects collected by the UAV.The algorithm can extract river of the river from the video frame image.The accuracy rate of ROI extraction is 83.47%.In order to further improve the recognition accuracy,the original algorithm add the rectangular similarity judgment of the river area,which improves the accuracy of the river extraction by 2.65%.At the same time,the algorithm add correction operation in case of trace extraction failure.The final river extraction accuracy rate is 92.31%.Secondly,the algorithm is used to extract river in the collected video.And the river area is saved as an image,while the image without river region is eliminated.Finally,the annotation software Label Image is used to annotate the saved image with floating objects,and the floating object dataset is generated.(2)Taking river video interception image as a sample,using soft data enhancement technology to expand the floating object image.The SSD and Faster R-CNN depth network of various feature extraction networks are used to train the sample and compare the results.The experimental results show that the Res Net-101-based SSD deep network recall rate is 61.67% and the F1 value is 71.29% with the accuracy of 84.47%,the Res Net-101-based Faster R-CNN deep network recall rate is 58.83% and the F1 value is 71.29% with the accuracy of 85.05%.Through the comparative analysis of experimental data,the SSD depth network based on Res Net-101 improves the accurate detection of river floating objects.The SSD and Faster R-CNN deep networks based on Mobile Net extraction network are used for sample training and the results are compared.The experimental results show that the accuracy of the Mobile Net-based SSD network model is 77.12% and the F1 value is 58.15%,the number of detected frames per second is 3.87 frames;the accuracy of the Mobile Net-based Faster R-CNN network model is 78.14% and the F1 value is 56.58%,the number is 2.67 frames Through the comparative analysis of experimental data,SSD deep network model recognition based on Mobile Net is faster than Mobile-based Faster R-CNN network model.(3)Designing and implementing an intelligent detection system for river floating objects.By designing the river area detection algorithm and analyzing the performance of the deep network model,using a variety of technologies applicable to the Android,a solution for a river floating intelligence detection system based on deep learning was put forward.