Applying a deep residual network coupling with transfer learning for recyclable waste sorting

MSWNet: A visual deep machine learning method adopting transfer learning based upon ResNet 50 for municipal solid waste sorting

“…Each bottleneck is consisted with the convolutional layer and batch normalization. Convolutions of three layers 1 × 1, 3 × 3 and 1 × 1 blocks in bottleneck 1 and bottleneck 2, where the function of the layer of 1 × 1 is reduced but the dimension of input is increased, making the layer of 3 × 3 a bottleneck with small input/output dimensions (Lin et al, 2022a). The function of average pooling layer is subsampling the pixels that would not change the object, and after that, the weights and bias of each neuron would be transmitted to fully connected layer.…”

“…This process was repeated until the input pixel space was achieved. In addition, the algorithm of PAC and t-SNE were also taken to improve the reasonable analysis of low-dimension MSW images, which can be seen in previous studies (Lin et al, 2022a).…”

“…In addition, feature maps of the training and validating dataset were also used to visualize the MSWNet vias deconvolution, PCA and t-SNE. While the performance of MSWNet in the test dataset was evaluated by many indices, which can be referred in previous studies (Lin et al, 2022a).…”

“…The mechanism of cyclical learning rate adopts the principle of triangular rate, as shown in Fig. S3, which can be referred in previous study (Lin et al, 2022a). The values of minimum bound and maximum bound of learning rate in the experiments was set as 10 −7 and 10, respectively.…”

“…In particular, organic waste was clearly separated from hazardous waste, recyclable waste and residual waste. While the feature of hazardous waste and residual waste was overlaped, this phenomenon can be explained that data from high dimension to two dimensions obtained by the method of PCA was used to calculate the eigenvector and eigenvalue of covariance matric (Lin et al, 2022a).…”