Few-shot Garbage Classification And Detection Method Based On Deep Learning

The classification and recycling of garbage not only helps to reduce the harm of garbage to the environment but also can produce huge economic benefits through the reuse of recycled garbage.However,garbage classification and recycling mainly rely on human eyes to classify garbage,which has the problems of low recycling efficiency and large consumption of human resources.Therefore,automatic garbage classification is the development trend in environmental protection.Vision-based automatic garbage classification is to judge the category and location of garbage through visual sensors and sort the garbage by mechanical devices.Automatic garbage classification can avoid falling into the dilemma of artificial garbage classification and improve recycling efficiency.In the fields of security,medical treatment,and automatic driving,the classification and detection based on deep learning have made significant progress.Therefore,applying deep learning to visionbased automatic garbage classification equipment to improve the speed and accuracy of garbage classification will have important research significance.Because there are many kinds of garbage and the appearance of the same kind changes significantly,it takes much workforce to collect and label garbage images and build large datasets.The number of existing public garbage classification and detection datasets is small,and the number of samples in the dataset is relatively insufficient,which brings difficulties to applying deep learning technology in a garbage classification system based on vision.Aiming at the garbage classification and detection method under the insufficient number of samples,we have carried out in-depth research in this thesis.The main research contents of this thesis are as follows:(1)To tackle the problem of the mutual occlusion and too small scale of objects,this thesis proposed a garbage detection algorithm based on adversarial spatial dropout module and feature fusion,named Cascade Adversarial Spatial Dropout Detection Network(ASDDN).This algorithm proposes an adversarial space dropout(ASD)module that can generate occlusion samples.In this module,the Grad-CAM algorithm is used to obtain the heat map reflecting the influence of different regions of the feature map on the classification results.After the heat map is transformed into a mask,it is combined with the feature map to obtain the occlusion sample.During model training,adding the dropout module to the detection network can improve the detection ability of the detection model for partially occluded garbage.At the same time,by adding the FPN network after the feature extraction network of the model for feature fusion,the semantic information of small-scale garbage on the feature map is enhanced,and the detection effect of the detection model on small-scale garbage is improved.(2)Aiming at the problem that the classification model trained by the traditional deep learning method cannot adapt to the garbage classification task with insufficient samples,a Meta-Res Net few-shot garbage classification algorithm based on MAML and improved Res Net is proposed.This algorithm improves the ability of the model to fit data by removing the maxpooling layer of the feature extraction part of the Res Net-18 network and adding the full connection layer of the classification part.At the same time,it reduces the impact of different data feature distributions on the model’s accuracy by introducing group normalization.In the mini-Image Net dataset,the improved classification model is trained by Model Agnostic Meta-Learning(MAML)method.When the trained model is used for the garbage classification task,it can converge quickly and obtain a good classification effect after training with a few garbage samples.(3)In the few-shot detection model based on Faster RCNN,the RPN network will generate false proposal boxes that do not belong to the supported sample category,and the relationship network of the FSOD model cannot effectively measure the distance between different categories,resulting in the weak recognition ability.Based on the FSOD model,this thesis proposes a few-shot garbage detection algorithm AT-FSOD based on attention mechanism and hard sample triplet loss function.Firstly,this algorithm proposes a new dual attention module to filter the feature information irrelevant to the category of support samples on the query sample feature maps so that the RPN network only generates proposal boxes close to the category of support samples.Then,the hard-triple loss function is used to increase the distance between different categories in the feature space and improve the classification ability of the model.Finally,the deformable convolution network is added to the feature extraction network to reduce the difference of feature information of the same kind of garbage and improve the detection effect of the detection model on deformable garbage.