Natural Image Object Detection Based On Lightweight Neural Networks

As one of the most basic problems in computer vision,object detection is the foundation for other complex vision tasks.In recent years,the rapid development of deep learning and computing resources has promoted breakthroughs in object detection and accuracy.However,in some practical scenarios,deep network structure with hundreds of layers usually contains a great quantity of parameters and floating-point calculations,which requires a large amount of storage and running space.It is often difficult to meet the requirements of real-time detection and deploy on resource-constrained terminal devices.In order to promote the application of object detection model from servers to mobile terminals,the models need to be light-weighted.The following researches are carried out in this paper:(1)According to the repeated gradient information during the network optimization,an improved lightweight YOLOv3 object detection method is proposed,which prevents repeated gradient information by truncating the gradient flow,enhances the learning ability of CNN,and achieves richer gradient combinations.The experimental results show this method achieves the reduction of computational complexity and parameters.Meanwhile,the detection speed is also improved without losing too much accuracy.(2)A detection method fused Transformer encoder blocks and convolutional neural network is proposed.Transformer can capture global dependencies and rich contextual information,and CNN has some visual inductive bias like locality and translation equivalence.Our method avoids the heavy computational burden caused by Transformer’s multi-head selfattention with long patches sequence.Experiments results show our methods improve detection accuracy and inference speed without adding too many parameters.(3)An object detection method based on model scaling is proposed,which further shrinks the model and balances the computational and memory cost of lightweight neural networks.The reason why Dense Net has fewer parameters than Res Net and consumes more time and energy is explored,and the factors of efficient network design is also considered.The experimental results show it has fewer model parameters and faster detection speed under the premise of a certain accuracy.