Improved garbage image classification for inception-Resnet-V2
Yuehua Dong, Peng Huilin · 2022 Asia Conference on Algorithms, Computing and Machine Learning (CACML) · 2022
In view of the current situation that the development of deep learning convolutional neural network models tends to be high in depth, the number of model parameters is large and the training time is continuously increasing. A structure based on the combination of Inception-ResNet-V2 and depthwise separable convolution is proposed, using depthwise separable volume. The integration separates channels and regions, thereby reducing a large number of parameters; at the same time, in order to extract more critical information in the feature map and improve the classification effect of the model, the attention mechanism CBAM module is introduced. In the experiment, the image of the garbage dataset is used as the experimental object, and the model is trained by data enhancement methods such as image cropping, rotation, and flipping of the original data. In order to verify the best performance of the improved model in the deep network, the experiment conducted model tests of different depths. The experimental results show that as the network continues to deepen, the classification effect of the model shows a trend of increasing first and then decreasing. When the number of Inception-ResNet-A, B, and C modules is 7,14, and 7, the model of the improved Inception ResN et- V2has the highest classification accuracy, reaching 95.80%, which is 4.58% higher than the Inception-Resnet-V2 classification accuracy, and has less parameters and better classification effect, which has higher practical value.