? Based on DenseNet's dense connection idea, through a series of structural optimization, a mobile device-oriented network structure PeleeNet is proposed, and a target detection network Pelee is proposed by integrating SSD. From the experimental point of view, PeleeNet and Pelee are good choices in speed and accuracy.
? Based on DenseNet's thought, PeleeNet added several key improvements.
? Inspired by GoogLeNet, the original dense layer is changed to two dense layers, as shown in figure1b. The new path superposes two convolutions to learn and obtain different receptive field features, especially large object features.
? DenseNet uses convolution with step size =2 to preliminarily process the input. Inspired by Inception-v4 and DSOD, this paper designs an efficient stemming block, whose structure is shown in Figure 2, and two paths can extract different features. This can improve the feature expression ability of the network without too much time-consuming calculation.
? In DenseNet, the bottleneck layer is used to compress the input elements, but the output of the bottleneck layer is fixed at four times that of the dense layer. In the early layer of the network, the output of the bottleneck layer will be greater than the input, which will lead to the decrease of efficiency. Therefore, this paper links the output scale of the bottleneck layer with the input scale to ensure that it is not greater than the input scale. As can be seen from Figure 4, the calculation efficiency is doubled after modification.
? DenseNet uses the transition layer between dense layers to compress the feature size, and the compression ratio is 0.5. Through experiments, it is found that this operation will weaken the expressive ability of network features, so the output dimension of the transition layer is fixed to the size of the input and is no longer compressed.
? DenseNet adopts the pre-activation mode of Conv-Relu-BN, which is modified to the post-activation mode of Conv-BN-Relu in this paper, which makes the convolution layer and BN layer combine in reasoning and improves the calculation speed. In addition, convolution is added to the final dense layer to obtain stronger feature expression ability.
? The structure of PeleeNet is shown in Table 1, which includes a stem block, four feature extraction stages and a pool layer with an average span of 2. In this paper, whether to use three feature extraction stages or four feature extraction stages is very tangled. These three stages need stem blocks to further reduce the size of the feature map. Considering that reducing the size of the feature map too quickly at first will weaken the expression ability of the network, four stages are adopted.
? On the basis of SSD, PeleeNet is used as the backbone network, and a target detection network Pelee is put forward after many optimizations. The main optimization points are as follows:
? Verify the key features of PeleeNet.
? Compared with other lightweight networks.
? Comparison of actual reasoning speed of PeleeNet on different devices.
? Pelee compares with other network settings.
? Performance comparison of various improvement measures.
? Compared with other networks in VOC performance.
? Comparison of reasoning speed of various devices.
? Performance comparison between COCO and other networks.
? Peleet is a variant of DenseNet, which does not use the popular deep integrable product. Peleet and Pelee have achieved very good performance and speed only through structural optimization. After reading the paper, they can learn many skills of network design.
?
?
?