Expression Recognition CNN v3.0.0

中文

V3.0.0 将普通卷积层级升级为手动实现的轻量级残差网络。每个残差块包含两层 3×3 卷积、Batch Normalization 和 shortcut connection，不使用预训练 ResNet。

三个独立残差模型的 combined FER2013 test accuracy 分别达到 63.90%、63.33% 和 63.65%。最终版本通过等权 soft voting 平均三个模型的 Softmax 概率，将整体准确率进一步提升至 66.59%。

该版本同时提供独立 Windows 程序 ExpressionRecognitionCNN-v3.0.0.exe，支持图片选择、预览和 Top-3 表情概率输出。

English

V3.0.0 replaces the plain convolutional stages with a hand-built lightweight residual network. Each residual block uses two 3×3 convolutions, Batch Normalization, and a shortcut connection, without relying on a pretrained ResNet.

The three independently trained residual models achieve 63.90%, 63.33%, and 63.65% accuracy on the combined FER2013 test set. The final equal-weight soft-voting ensemble averages their Softmax probabilities and improves overall accuracy to 66.59%.

This version also provides a standalone Windows application, ExpressionRecognitionCNN-v3.0.0.exe, with image selection, preview, and Top-3 expression probability output.

Expression Recognition CNN v2.0.0

中文

V2.0.0 在原始四层 CNN 的基础上，为每个卷积层加入 Batch Normalization，并采用分阶段学习率衰减策略。模型在前 10 个 epoch 使用 1e-3 学习率，后 10 个 epoch 降至 3e-4。

该版本在 combined FER2013 test set 上的准确率从 V1.0.0 的 56.56% 提升至 61.87%，并明显改善了 angry、disgust、fear 和 sad 等原本表现较弱的类别。

English

V2.0.0 adds Batch Normalization after every convolutional layer and introduces staged learning-rate decay to the original four-layer CNN. The learning rate is reduced from 1e-3 to 3e-4 after the first 10 epochs.

Accuracy on the combined FER2013 test set improves from 56.56% in V1.0.0 to 61.87%, with substantial improvements in several previously weak classes, including angry, disgust, fear, and sad.