Research On Garbage Classification Based On Conditional Confrontation Network

The current process of garbage disposal usually involves residents throwing domestic garbage into trash cans,sanitation workers cleaning it,and then transporting it to a designated treatment plant for incineration or landfill.This will lead to insufficient use of part of the recyclable resources,and at the same time cause huge pollution to the environment.However,what residents throw away is to classify the garbage images and put the corresponding garbage into the corresponding garbage bins,so that useful garbage can be put forward for use.The advantages of this treatment are:resources are recycled and environmental pollution is reduced,Labor cost reduction,etc.The garbage images are classified into the garbage,and the types of garbage are fixed,such as recyclable garbage,kitchen waste,hazardous garbage,etc.But the image of the same kind of garbage may have many different states,including whole oranges that have fallen to the ground,and also oranges that have been rotten in half.Often when designing and training a network model,there are not enough samples to train the network model,which will result in insufficient model training to achieve the desired result we ultimately want.Traditional deep learning networks in the field of image recognition will have problems such as unstable training effects or insufficient image recognition accuracy.This paper proposes an improved image recognition method based on a conditional generation confrontation network for the above problems.First of all,based on the conditional generation confrontation network,we use the advantages of the generator to add category labels to control image generation and use the images generated by the generator as training data to achieve the purpose of expanding the data set.At the same time,add a deconvolution layer and a convolution layer in the generator and discriminator to optimize the network model,and then remove the fully connected layer to ensure the improvement of model stability.Introduce conditional batch normalization,use it to retain the advantages of category labels,so that the network model can learn more image features.Finally,the improved discriminator is a classifier,and a new objective loss function is proposed,the model convergence speed is improved,and the recognition accuracy can be improved to a certain extent.