| 摘 要: 提出轻量模型Mini Net用于实时检测，并保证其准确度。Mini Lower利用Group卷积与通道合并提取低阶特微，Mini Higher利用可分离的Depthwise卷积提取高阶特微。Mini模块实现的高效卷积使其大幅减少了参数量与计算量，并且在空间维度上引入更多层次所带来的非线性，可提升模块的提取能力。另外，在模型中利用更精细的特微搭配多尺度预测改善小目标检测。基于一系列的消融实验验证Mini模块设计的有效性，并透过对照实验结果证实MiniNet模型的实时性优于全卷积模型，在参数量仅有0.92×106的情况下，能够有效地提取目标特微。 |
| 关键词: 卷积神经网络；轻量模型；目标检测；图像识别 |
|
中图分类号: TP311
文献标识码: A
|
| 基金项目: 天津市大学生创新创业训练计划项目(202110069073). |
|
| Research on Convolutional Neural Networks in Real-time Detection |
|
GAO Xinyi, CHEN Qi, CHEN Guanyu, YANG Jingyi, ZHANG Kunkun,CAI Huarui
|
(Tianjin University of Commerce, Tianjin 300134, China)
1275475886@qq.com; chq687002@163.com; 2920734342@qq.com; 2385333215@qq.com; 2529207862@qq.com; 1431502854@qq.com
|
| Abstract: This paper proposes a lightweight model Mini Net for real-time detection and its accuracy is guaranteed. Mini Lower uses Group convolution and channel merging to extract low-order micros, while Mini Higher uses separable Depthwise convolutions to extract high-order micros. The efficient convolution implemented by the Mini module greatly reduces the amount of parameters and computation, and the nonlinearity brought by more layers in the space dimension is introduced, which can improve the extracting ability of the module. In addition, a combination of a finer micro and multi-scale prediction is used in the model to improve small object detection. Based on a series of ablation experiments, the effectiveness of the Mini module design is verified, and the comparative experimental results very that the real-time performance of the Mini Net model is better than that of the full convolution model. When the parameter amount is only 0.92×106, the target micro can be extracted effectively. |
| Keywords: convolutional neural network; lightweight model; object detection; image recognition |
