Convolutional Neural Network-Based Multiscale Feature Selection and Evaluation in Image Segmentation

Multiscale image segmentation based on artificial neural networks is a hot topic in research on remote sensing image processing. However, the establishment and evaluation of pooling models and selection of feature operators lack clear standards. Based on the biological visual multiscale perception mechanism, this study combines classical wavelet theory with convolutional neural network theory to establish 10 sets of geometric operators and construct the corresponding multiscale image feature pyramids. Statistical analysis shows that the 10 sets of operators exhibit two types of information transmission characteristics, that is, balanced and growth. The obtained image features become more fragmented as operator complexity increases. After excluding the two operator groups with high complexities, the remaining eight groups were applied to the convolutional neural network image-segmentation algorithm. Eight pooling models were established to obtain the corresponding multiscale image features, perform convolution operations, and generate multiscale segmentation results for remote sensing images. The evaluation results reveal that the high complexity of the feature operators is unfavorable for feature transmission and preservation, and compared with operators having the information transmission characteristics of growth, those with balanced information transmission characteristics show better performance in convolutional neural network image segmentation. The segmentation accuracy was improved by 1.5%–2%. The conformity of the segmentation results was improved by 1%–1.5%. Finally, the degree of interclass chaos is reduced by 4.1%–10%.