Convolutional Neural Networks with Modified Evaluation Functions for Image Semantic Segmentation

Image semantic segmentation (ISS) is a classic problem in the field of image technology, and artificial intelligence-based neural network segmentation methods are widely used in ISS. However, as the accuracy gradually converges to higher levels, the computational overhead from complex models becomes increasingly difficult to ignore. This paper focuses on this problem in practical application, and proposes the use of deep separable convolutions in restricted computational resource scenarios with a large number of model parameters, and the optimization of the original cross-entropy loss function for the loss function of small-sample segmentations. The experimental results show that model selection in deep learning tasks is dataset-oriented, and small-sample learning optimization is effective in improving a model's a bility to perceive details.