Research And Application Of Object Detection Algorithm Based On Lightweight Deep Convolutional Neural Network

Object detection is the process of finding the position and category of an object of interest from an image.With the continuous development of deep learning,the accuracy of object detection is constantly refreshed.Higher precision object detection algorithm can be applied to more field.The object detection based on deep convolutional neural network can extract deeper semantics through multi-layer networks.By improving the feature components of feature extraction,the accuracy and recall of the model can be improved.In the recent development of object detection algorithms,anchor-free network occupies the research hotspot.Because the anchor-free detection algorithm can remove the process of post-processing NMS and replace it in the whole network,it is more efficient than the traditional algorithm and requires fewer hyperparameters to be set.This paper is based on the anchor-free object detection network algorithm Center Net(Objects as Points)for improvement and optimization.We mainly optimize the pre-processing,backbone network,up-sampling and loss function of the current algorithm.Optimize the network with a lightweight,small backbone network.The accuracy of the model is improved without affecting the reasoning speed too much.The main work contents of this paper are as follows:1.Optimize the pre-processing stage by adding multi-scales to the model in training stage.2.Optimize the backbone network.The lightweight convolutional neural network Mobile Net V2 backbone network is used,and the network is optimized by adding attention module and feature fusion weight.The characteristics of channels are analyzed and more features that can be used are excavated.In the case of the use of channel mean feature information and maximum feature information,we also added the mean feature information of each channel to better express the data distribution of each layer feature map.Different neuronal compression ratios were tested to reduce the neuron shrinkage ratio of the channel attention mechanism,so that the cost performance of the attention module was higher.3.Optimize the upsampling process.We add a horizontal connection composed of3*3 deformable convolution to the network to make full use of more low-level semantic features.The feature map generated by horizontal connection and the up-sampled feature map use the fusion feature.The multi-layer subsampled feature images are combined with the size of output feature images by bilinear interpolation,so that the network not only has sufficient high-level semantics,but also complements the low-level semantics.4.Modify the loss function of the network,introduce the Io U loss function to optimize the problem that the detection of small and medium-sized objects is not good enough,and improve the border regression quality of small and medium-sized objects,so as to improve the recall rate.