Research on Intelligent Video Detection of Small Targets Based on Deep Learning Intelligent Algorithm
Table 1
Brief introduction and comparison of lightweight networks.
Model name
Reference volume/106
Basic introduction
Classification effect on ImageNet/%
MobileNet v1
4.20
Lightweight network model that can be used for mobile, using deep separable convolution instead of normal convolution; the number of parameters and the amount of computation are greatly reduced, but the straight-cylinder structure is not sufficient for feature learning
70.6
MobileNet v2
3.40
The hourglass residual structure is introduced to enhance the gradient propagation and reduce the computation, and the ReLU function is removed from the last layer to preserve the feature diversity
72.0
MobileNet v3
5.40
The model structure is improved by using the network structure search algorithm, while the SE module is introduced to strengthen the network learning ability by combining with the channel attention mechanism, and the h-swish activation function is proposed to improve the accuracy
75.2
ShuffleNet v1
1.90
The point-by-point group convolution is used to reduce the computational complexity of 1 × 1 convolution, and the channel transformation method is proposed to induce the flow of information in the same channel of different features, but the difference between the number of input and output channels is still too large to affect the efficiency
67.8
ShuffleNet v2
2.30
Abandon group convolution and introduce channel splitting operation to reduce the number of network branches to obtain faster detection speed with certain accuracy
69.4
ShuffleNet
1.25
Replace the 3 × 3 convolution with 1 × 1 convolution, reduce the number of convolution channels, and postpone the sampling operation to significantly reduce the number of parameters and the amount of computation so as to maximize the speed increase in exchange for the reduction in accuracy
