Recyclable Garbage Detection And Applications Based On Improved YOLOv5s

There are many benefits of garbage recycling,which can help to protect water and soil resources,improve the quality of residents’ living environment,and accelerate the development of green circular economy.However,traditional garbage recycling requires a lot of manpower and material resources.In order to obtain higher accuracy,target detection models usually have a large number of parameters and complex structures.This study presents a lighter and more efficient YOLOv5 s improved model for the classification and positioning of recyclable garbage,and practices related to the application development of recyclable garbage detection.The main work is as follows:1.For the original YOLOv5 s network,the problem of large number of parameters and complex calculation part,is improved by properly combining Shuffle Net V2 basic unit,deep separable convolution and Hard-Swish activation function,which reduces the redundancy of the network and makes the structure more compact.2.In the original YOLOv5 s network,the network can distinguish the significance of different channels by embedding a lightweight SE attention mechanism,which gives different weights to the structure with more channels in feature mapping.3.Using genetic algorithm and K-means++ clustering algorithm,this work can get more accurate recyclable garbage anchor frame again,which is easy to adjust the network border regression.4.In order to improve the performance of the garbage collection detection model,the transfer learning method is used to migrate the selected dataset to COCO dataset,which enables it to learn the rich features of 80 categories in advance.5.For the actual embedded application development of the recyclable garbage detection model,the common Jenson Nano and Raspberry Pi4 B are selected as the micro-embedded platforms,and the actual running speed is compared and investigated.In order to facilitate the user’s operation,observation effect and adjusting threshold,a graphical interface is created by using Pyqt5 library to achieve the convenient operation of recyclable garbage detection.6.In order to research and implement the task of recyclable garbage detection on the network,a WEB platform for Collectible garbage detection is built using Flask back-end and VUE front-end.The experimental results show that the improved model parameters are compressed to 60.3% of the original model parameters and 59.2% of the model memory.At input resolution 640 × 640,the m AP of the improved model is 96.43%,which is 4.34% higher than the original YOLOv5 s.In terms of speed,the improved model improves the parallelism on GPU platform.By deploying on Jetson Nano hardware,the forward inference speed of the improved model is 13.1%faster than that of the original YOLOv5 s.In addition,compared with the current mainstream object detection models,the proposed improved model also has a good expression ability of recyclable garbage characteristics,which can provide reference for the lightweight development of recyclable garbage detection.