Objective

2

To optimize the detection accuracy， lightweight target detection method is focused on the problem of cost efficiency of computation and storage. The Mobilenetv3 can extract feature effectively through the inverted residual structures-related bneck. However， features are connected with bneck-inner only excluded the bneck-between feature connection. The network accuracy is not optimized because more initial features are not involved in. To achieve a better balance between computation and detection accuracy， the neural architecture search-feature pyramid networks lite（NAS-FPNLite） can be as an effective target detection method based on deep learning technique. The NAS-FPNLite detector is focused on depth separation convolution in the feature pyramid part and the channel number of the intermediate feature layer can be compressed to a fixed 64-dimensional. A better balance can be achieved between the floating point operations and detection accuracy. The depth separation convolution can configure the parameters to 0 easily under this circumstance. To resolve these two problems， we develop a NAS-FPNLite lightweight object detection method in terms of the fusion of cross stage connection and inverted residual. First， an improved network model CSCMobileNetv3 is illustrated， which can obtain more multifaceted information and improves the efficiency of network feature extraction. Next， the inverted residual structure is applied to the feature pyramid part of NAS-FPNLite to obtain a higher number of channels during the depthwise separable convolutions， which can improve the detection accuracy on possibility-alleviated of those parameters become 0. Finally， the experiment for CSCMobilenetv3 model is validated on the Canadian Institute for Advanced Research（CIFAR）-100 and ImageNet 1000 datasets， as well as the inverted residual NAS-FPNLite detector in the COCO （common objects in context） dataset.

Method

2

At the beginning， to obtain different gradient information between network layers， a cross stage connection （CSC） structure is proposed in terms of DenseNet dense connection， which can combine the initial input with final output of the same level network block to obtain the gradient combination with the maximum difference， and get an improved CSCMobileNetv3 network model. The CSCMobileNetv3 network model is composed of 6 block structures， the first two blocks remain unchanged， and the last four block structures are combined with the CSC structure. Within the same block， the initial input is combined with the final output and as the input of the next block. At the same time， to obtain more different gradient information， the number of channels between the various blocks is changed from the original 16， 24， 40， 80， 112， 160 to 16， 24， 40， 80， 160， 320 in correspondent， It can surpress the intensity of the number of parameters and the amount of floating point operations effectively derived from the excessive expansion of channels. Then， in the detector part of NAS-FPNLite， the feature pyramid part is fused with the inverted residual structure， and the feature fusion method of element-by-element growth between different feature layers is replaced by the channel-concatenated. It is possible to maintain a higher number of channels via processing depthwise separable convolutions. To perform sufficient feature fusion， the situation where the parameters become 0 is avoided effectively， and a skip connection is realized between the input feature layer and the final output layer. A NAS-FPNLite object detection method fused with inverted residuals is demonstrated.

Result

2

In the training stage， our configuration is equipped with listed below： 1） the graphics card used is NVIDIA GeForce GTX 2070 Super， 2） 8 GB video memory， 3） CUDA version is CUDA10.0， and 4） the CPU is 8-core AMD Ryzen7 3700x. On

CIFAR-100 dataset training， since the image resolution of the CIFAR-100 dataset is 32 × 32 pixels， and the CSCMobileNetv3 is required of the resolution of the input image to be set to 224 × 224 pixels， the first convolutional layer， and in the first and third bnecks， the convolution stride is set from 2 to 1， 200 epochs are trained， the learning rate is set to multi-stage adjustment， the initial learning rate lr is set to 0.1， and then at 100， 150 and 180 epochs， multiply the learning rate by a factor of 10. On ImageNet 1000 dataset training， the dataset needs to be preprocessed first， the image resolution is adjusted to 224 × 224 pixels as the input of the network， the number of training iterations is 150 epochs， the cosine annealing learning strategy is used， and the initial learning rate lr is 0.02. The experimental results show that the accuracy of the CSCMobileNetv3 network can be increased by 0.71% to 1.04% compared to MobileNetv3 when the scaling factors are 0.5， 0.75， and 1.0 on the CIFAR-100 dataset. Compared to MobileNetv3， CSCMobileNetv3 increases the number of parameters by 6% and the amount of floating point operations by about 11% when the scaling factor is 1.0， but the accuracy rate increases by 1.04%. Especially， when the CSCMobileNetv3 zoom factor is 0.75， the number of parameters is reduced by 30% compared to the MobileNetv3 zoom factor of 1.0 and the amount of floating point operations is reduced by 20%. The accuracy rate is still improved by 0.19%. On the ImageNet 1000 dataset， CSCMobileNetv3 has a 0.7% improvement in accuracy although the amount of parameters and floating point operations is slightly higher than that of MobileNetv3. To sum up， CSCMobileNetv3 has optimized relevant parameters， floating-point operations and accuracy. On the COCO dataset， compared to other lightweight object detection methods， the detection accuracy of CSCMobileNetv3 and NAS-FPNLite lightweight object detection method fused with inverted residuals can be improved by 0.7% to 4% based on equivalent computation.

Conclusion

2

The CSCMobileNetv3 can obtain differential gradient information effectively， and achieve higher accuracy with a small calculation only. To improve detection accuracy， the NAS-FPNLite object detection method fused with inverted residuals can avoid the situation effectively where the parameters become 0. Our method has its balancing potentials between the amount of calculation and detection accuracy.