Intelligent Recognition Of Domestic Garbage Based On Object Detection

With an increased municipal waste production per capita,the efficiency of the waste treatment system needs to be improved to consume and recycle the ever-increasing amount of municipal waste.Image intelligent recognition based on convolutional neural network can be trained to capture image features for detection according to the task setting,which is very suitable for processing large number of pictures with complex categories,and is expected to be an important detection method in the waste treatment system.Considering the limitation of the current waste classification and sorting technology and the systematic of the waste treatment process,it is of great significance to introduce new technologies and integrate them into the entire waste treatment system for assessment.This article first analyzes the characteristics of image intelligent recognition and the waste treatment system,and obtains the specific implementation methods of image intelligent recognition technology applied on each aspect of the waste treatment system.Among them,image classification technology is mainly used for source garbage classification,and object detection technology is mainly used for garbage sorting and garbage intelligent monitoring.Object detection algorithm based on convolutional neural network can accurately identify and locate multiple targets with complex categories,which is more adaptable and is available to improve the disadvantages of the existing garbage classification and sorting technologies.However,the current research on object detection for waste processing is not sufficient and has some typical drawbacks such as low dataset quality,low combination with garbage processing characteristics and insufficient consideration on generalization performance.Therefore,the experimental part of this paper establishes a garbage dataset and comprehensively studies the application of a typical object detection algorithm named Faster RCNN on garbage detection.The comparative analysis is carried out with three types of backbone networks with significant differences,including VGG-16,Res101 and Mobile Net＿v1.Typical 6 kinds of garbage with 6076 images were collected from recyclable and hazardous waste,and were labeled to create the target dataset with the semi-automatic annotation method which was developed to reduce the labeling workload.The data augmentation method is used to expand the targets quantity and the targets scale diversity,which makes the dataset more balanced and facilitate algorithm training and testing.The research uses end-to-end training combined with the special layer fine-tuning for network training to well balance the precision and speed requirement,and carries out enhanced training on low recognition rate samples to obtain a minimum mean average precision of 92.85%.Subsequently,three typical error are captured and analyzed from the misidentified samples,which was used to get the optimization method and increase the highest recognition mean average precision to 99.23%.The setting method of probability thresh for object detection is analyzed and obtained,which can raise the precision of recycled waste and the recall of the hazardous waste.In addition,in order to analyze the generalization performance of different backbone network embedded in the algorithm,a dataset with 816 pictures derivatized from different background was built and used to test the impact of changing background on garbage detection.It was found that the complex backgrounds from surrounding garbage put the greatest impact on detection accuracy,and the generalization performance has the result of Res101 > VGG-16 > Mobile Net＿v1.Finally,the overall process of the image intelligent sorting system is designed.