Object detection method and electronic apparatus
View Patent ↗An object detection method and an electronic apparatus are provided. A processor is configured to perform the following. An original image is received. A plurality of initial feature layers of different scales is extracted from the original image. A plurality of fused feature layers of different scales is obtained through an interpolation and an addition operation based on the initial feature layers. The fused feature layers are respectively input into corresponding detection heads to obtain a bounding box location probability distribution through a bounding box regression branch of the detection head and obtain a classification probability distribution through a classification branch of the detection head.
1. An object detection method implemented by a processor configured to execute an object detection module, the method comprising:
receiving an original image;
extracting a plurality of initial feature layers of different scales from the original image through a sampling path, wherein the sampling path comprises a plurality of sampling steps, wherein an initial feature layer obtained from each of the plurality of sampling steps is taken as an input of next of the each of the plurality of sampling steps;
obtaining a plurality of fused feature layers of different scales through an interpolation and an addition operation based on the initial feature layers through a fusion path, wherein the fusion path comprises a plurality of fusion steps, wherein a fused feature layer obtained from one of the plurality of fusion steps and an initial feature layer obtained from next of the one of the plurality of fusion steps are taken as an input of the next of the one of the plurality of fusion steps, wherein each of the plurality of fusion steps comprises:
performing the interpolation on the fused feature layer taken as the input to adjust a size of the fused feature layer taken as the input to be the same as a size of the initial feature layer taken as the input; and
performing the addition operation on the fused feature layer after adjusting the size and the initial feature layer to obtain a fused feature layer corresponding to the each of the plurality of fusion steps; and
respectively inputting the fused feature layers into corresponding detection heads to obtain a bounding box location probability distribution through a bounding box regression branch of the detection head and obtain a classification probability distribution through a classification branch of the detection head.
2. The object detection method according to claim 1 , wherein extracting the initial feature layers from the original image comprises:
extracting the initial feature layers of different scales from the original image by downsampling.
3. The object detection method according to claim 1 , wherein obtaining the fused feature layers of different scales comprises:
performing the interpolation and the addition operation based on the initial feature layers from a deep layer to a shallow layer by upsampling, and obtaining the fused feature layers.
4. The object detection method according to claim 1 , wherein after obtaining the fused feature layers, the method further comprises:
performing the interpolation and the addition operation based on the fused feature layers from a shallow layer to a deep layer by downsampling, and obtaining a plurality of new fused feature layers to respectively input the new fused feature layers into the corresponding detection heads.
5. The object detection method according to claim 1 , wherein the object detection module comprises a backbone, a neck, and the detection head.
6. The object detection method according to claim 5 , wherein a lightweight neural network is adopted for the backbone, and the lightweight neural network comprises an EfficientNet-Lite algorithm.
7. The object detection method according to claim 5 , wherein a path aggregation network is adopted for the neck.
8. The object detection method according to claim 1 , wherein a Generalized Focal Loss V2 algorithm is adopted the detection head.
9. The object detection method according to claim 1 , wherein knowledge distillation is introduced into the bounding box regression branch.
10. An electronic apparatus comprising:
a storage device comprising an object detection module; and
a processor coupled to the storage device, wherein the processor is configured to input an original image into the object detection module to:
extract a plurality of initial feature layers of different scales from the original image through a sampling path, wherein the sampling path comprises a plurality of sampling steps, wherein an initial feature layer obtained from each of the plurality of sampling steps is taken as an input of next of the each of the plurality of sampling steps;
obtain a plurality of fused feature layers of different scales through an interpolation and an addition operation based on the initial feature layers through a fusion path, wherein the fusion path comprises a plurality of fusion steps, wherein a fused feature layer obtained from one of the plurality of fusion steps and an initial feature layer obtained from next of the one of the plurality of fusion steps are taken as an input of the next of the one of the plurality of fusion steps, wherein each of the plurality of fusion steps comprises:
performing the interpolation on the fused feature layer taken as the input to adjust a size of the fused feature layer taken as the input to be the same as a size of the initial feature layer taken as the input; and
performing the addition operation on the fused feature layer after adjusting the size and the initial feature layer to obtain a fused feature layer corresponding to the each of the plurality of fusion steps; and
respectively input the fused feature layers into corresponding detection heads to obtain a bounding box location probability distribution through a bounding box regression branch of the detection head and obtain a classification probability distribution through a classification branch of the detection head.