Fig. 11

The data that are incorrectly predicted by the serial models such as ResNet101, ResNet152, ResNet161, and ResNet201, but correctly predicted by the parallel models. a denotes unprocessed data, b, c, d are feature maps generated by HAHNet convolution. The convolution order starts with the convolution in (b), followed by (c), and finally (d). The convolutions in b, c, and d each consist of a single convolutional layer and are repeated only once. We extract and display the feature maps corresponding to four channels from each convolution. The highlighted part of the feature map is the feature noticed by the model