The emergence of ResNet provides a powerful tool for training extremely deep networks. The core idea behind it is to change the learning goals of the network. It no longer learns new features from scratch but learns the difference between the target and existing features. However, the difference between the two kinds of features does not have an independent and clear meaning, and the amount of learning is based on the absolute rather than the relative difference, which is sensitive to the size of existing features. We propose a new network that perfectly solves these two problems while still having the advantages of ResNet. Specifically, it chooses to learn the quotient of the target features with the existing features, so we call it the quotient network. In order to enable this network to learn successfully and achieve higher performance, we propose some design rules for this network so that it can be trained efficiently and achieve better performance than ResNet. Experiments on the CIFAR10, CIFAR100, and SVHN datasets prove that this network can stably achieve considerable improvements over ResNet by simply making tiny corresponding changes to the original ResNet network without adding new parameters.

Wheat head detection is a critical component in high-throughput phenotyping, holding significant application value for wheat yield estimation and breeding analysis. With the continuous advancement of general object detection models, state-of-the-art detectors achieve high accuracy in same-distribution wheat head detection scenarios. However, when applied to cross-distribution environments, their performance often degrades significantly. To investigate this issue, this paper adopts the official division of the GWHD2021 public dataset as the cross-distribution evaluation setting. Using the recently state-of-the-art object detection model YOLOv13 as the baseline, we systematically explore effective approaches to enhance cross-distribution generalization performance in wheat head detection. Specifically, we propose two lightweight modifications to YOLOv13’s Full-PAD architecture and HyperACE’s core modules, analyzing their potential mechanisms: (1) Replacing scalar gating in Full-PAD with channel-level gating enables finer-grained branch injection control. Concurrently, channel-level gating introduces equivalent stronger weight penalties during training, generating additional regularization effects. Through exploratory controlled experiments aligning weight-penalty strengths, we find that this gain depends on both channel decoupling and the accompanying implicit regularization rather than on decoupling alone. (2) Removing Batch Normalization from HyperACE’s core C3AH modules and adopting normalization strategies independent of batch statistics—such as Identity, Group Normalization, or Instance Normalization. Preliminary experiments show that while this results in a small, directionally positive change, the change remains within run-to-run variance; we therefore do not claim BN removal as a reliable standalone improvement. Furthermore, combining channel gating with BN removal does not yield further additive improvements compared to channel gating alone, indicating an interaction effect. To address this, we analyze relevant statistics between channel gating and HyperACE branches, providing an exploratory explanation for this non-additive phenomenon. In summary, this paper delivers empirical evidence and preliminary insights for enhancing generalization performance in the cross-distribution wheat head detection task of the advanced general-purpose detection model YOLOv13.