Research On Domestic Waste Detection And Classification Based On YOLO Model And Lightweight Algorithm

Nowadays,the amount of domestic waste generated by people increases exponentially with the improvement of living standards and consumption levels.A large number of cities are facing the problem of "garbage besieged",which seriously affects the beauty of the city.At present,the classification and recycling of municipal solid waste mainly depends on manual sorting,which is extremely inefficient and has a high error rate.In recent years,AI technology has rapid application and development,it has begun to empower all walks of life.Therefore,it is possible to use AI application technology to classify and recycle household waste.At the heart of AI is deep learning,which has inherent advantages in object detection.Firstly,introduces the latest research progress of object detection technology in garbage classification.At the same time,in view of the huge amount of parameters and computation of the classic YOLO v3 model,the inference speed is not fast,and it is inconvenient to deploy to mobile devices,three solutions are proposed in three different stages.In the first stage,an improved lightweight backbone feature extraction network Mobile Net v2-SE is used to replace the original Dark Net53.In the second stage,a fusion feature mechanism that can improve performance is added to the detection network,and the classification part and prediction box regression part of the loss are replaced,and the model is trimmed during training.In the third-stage engineering optimization,the improved operator fusion strategy and model quantization strategy based on the NCNN framework are used,and the final data type of the model is INT8.In order to reflect that the improved version of the target detection model also has a certain value in the field of engineering applications,an Android APP application is developed in the fifth chapter of the article,and the improved model is deployed to the mobile phone,and the garbage object detection is realized on the mobile terminal.The experimental data shows that the improved backbone feature extraction Mobile Net v2＿SE has an accuracy of 96.2% on the classification dataset.Dark Net53 has 1.9% higher accuracy,but 24 times as much computation and 7 times as much parameters.The m AP of the improved L-YOLO model on the garbage detection dataset is 87.04,and the model size is only 34 M.Finally,the size of the Q-YOLO model after engineering optimization is only 8.5M,and the m AP is 86.75.Compared with the YOLO v3,the average detection accuracy of Q-YOLO is 3.61 higher,and the model size is only 1/27,and the inference speed reaches 19 FPS,which basically achieves weak real-time performance.