Waste plastic image segmentation method based on dense connection

The invention discloses a waste plastic image segmentation method based on dense connection, and the method comprises the following steps: collecting a waste plastic image, obtaining a collected image, dividing the collected image into a verification set and a training set, and carrying out the semantic label marking of the collected image through an image marking tool; carrying out local data augmentation, preprocessing the augmented local image, converting the local image and the corresponding semantic label into tensors, and carrying out regularization on the image tensors; conducting modeltraining: inputting the training data into a semantic segmentation model based on dense connection, evaluating a training result by utilizing a cross entropy loss function with a focusing parameter and a balance factor, and performing back propagation by utilizing an Adam algorithm to update model parameters; and circularly training the model until a stop condition is reached, and storing the model and the model paramete