Research on deep learning image recognition technology in

Research on deep learning image recognition technology in garbage classification Qiang Guo* School of Software Engineering Faculty of Information Technology Beijing University of Technology Beijing, China *Corresponding author email: qbq 2013@emails. bjut. edu. cn Yuliang Shi School of Software Engineering Faculty of Information Technology Beijing University of Technology Beijing, China email: shiyl@bjut. edu. cn Shikai Wang School of Software Engineering Faculty of Information Technology Beijing University of Technology Beijing, China email: skwang@emails. bjut. edu. cn Abstract Nowadays, with the rapid development of industry, people's daily life is becoming more and more rich and diverse. However, while enjoying the pleasure brought by material life, a large number of garbage are produced. They are various, and people lack the knowledge related to garbage classification, which leads to the difficulty of manual or robot classification. This paper studies a garbage classification algorithm model based on deep learning convolutional neural network Efficientnet to help identify garbage classification. In this research, data augmentation and normalization are carried out to solve the problem of small amount of data sets and different sizes of pictures. Efficientnet is used to extract the features of images. In order to solve the problem that BN has no obvious effect on small batches in the network, we replace BN with group normalization (GN). In order to prevent some irrelevant information in the image from affecting the training of the model, we add attention mechanism after the output of Efficientnet to emphasize or select the important information of the target processing object, and suppress some irrelevant details, so that the model can focus on the key features and better identify the image; according to the above process, we use softmax to classify the spam image and divide it into four categories (Recyclables, Kitchen garbage , Hazardous garbage , Other garbage) The results show that the model can effectively extract the features of the input garbage image, and get accurate judgment, and identify the types of garbage. The experimental results show that the average accuracy of the algorithm model is high, and has good classification performance and robustness. In the practical significance of the research, this reliable model can help people quickly know the type of garbage, or can be applied to robot sorting, to help detect the types of garbage for robot judgment and sorting, so it has very important application scenarios and significance. Text With the rapid development of industry, people's daily life has become rich and colorful. A large number of daily necessities are used and consumed every day, which inevitably produces a large number of domestic garbage and industrial garbage. These garbage types are complex, difficult to distinguish, and difficult to sort, timeconsuming and labor-consuming. Therefore, when dealing with these garbage problems, some places will bury the garbage Incineration, these means it will cause great harm to the environment, and the harmful substances produced in this process will also endanger people's health. This method is not suitable only to solve and eliminate the garbage. We need to classify the garbage accurately, and then treat different kinds of garbage separately, which can effectively avoid the garbage to the environment and people body damage. However, from the current efficiency of garbage classification and treatment, people's cognition of the type of garbage is very poor. When putting the garbage into the garbage can of the corresponding type, it is often miscast. When the robot in the garbage treatment plant carries out garbage sorting, it first identifies the garbage, but the recognition accuracy is not satisfactory. Garbage accumulation brings difficulties in sorting, which can't classify garbage quickly and effectively. Therefore, our research on garbage identification and classification algorithm has very important practical significance, which can effectively assist people and robots to identify garbage classification. Figure. 1. Experimental process Figure. 2. Data Augmentation examples Figure. 3. Line chart of accuracy When we preprocess the data in the experiment, we use the way of data augmentation to prevent the over fitting problem. Data Augmentation is the random rotation, clipping, color jitter, horizontal and vertical flipping of images, which increases the amount of data set. Convolution neural network itself is invariant to the displacement, view angle, size and illumination of objects, that is to say, data augmentation does not affect the training of neural network, but will enhance the generalization ability of the model and improve the accuracy rate. Figure 2 shows an example of data augmentation. In this experiment, 40 kinds of garbage images were identified and predicted based on hazardous garbage, kitchen garbage, recyclable materials and other garbage categories. After 30 epoch training, the average accuracy reached 93. 47%, and the highest was 98. 3%. Figure 3 shows the broken line chart of accuracy rate of training set and test set with the increasing of epoch, and Figure 4 shows the broken line chart of loss rate, in which the loss rate is about 0. 3. Conclusion In view of the difficulty of garbage classification, this research uses convolutional neural network model in deep learning to solve the problem effectively. At the same time, we improve the network model and optimize the training process. The experimental results achieved the expected goal, the recognition accuracy is satisfactory, and the garbage classification is more effective very good promotion and improvement. This research can be applied to many application scenarios of garbage classification, such as intelligent garbage can, garbage sorting by robot in garbage treatment plant, garbage identification by mobile app, etc. , which is efficient and convenient, and has practical significance and broad application prospects. However, limited to the laboratory hardware conditions, our experiment may take a little longer, and the accuracy rate may be improved. We will continue to solve these problems in the future. In addition, we need to improve and optimize the model better and fully, strive to improve the recognition rate, and expand it in the later stage, such as being able to learn independently and identify others. This is our next research direction.