admin管理员组

文章数量:1794759

YoloV8改进策略:UniRepLKNet,大核卷积的最新成果,轻量高效的首选(全网首发)

摘要

将UniRepLKNet应用到YoloV8的改进中,经过测试,涨点明显,运算量也有下降!

代码语言:javascript代码运行次数:0运行复制
.2014.3001.5501
代码语言:javascript代码运行次数:0运行复制
br
代码语言:javascript代码运行次数:0运行复制
YOLOv8l summary (fused): 268 layers, 43631280 parameters, 0 gradients, 165.0 GFLOPs
                 Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 29/29 [
                   all        230       1412      0.922      0.957      0.986      0.737
                   c17        230        131      0.973      0.992      0.995      0.825
                    c5        230         68      0.945          1      0.995      0.836
            helicopter        230         43       0.96      0.907      0.951      0.607
                  c130        230         85      0.984          1      0.995      0.655
                   f16        230         57      0.955      0.965      0.985      0.669
                    b2        230          2      0.704          1      0.995      0.722
                 other        230         86      0.903      0.942      0.963      0.534
                   b52        230         70       0.96      0.971      0.978      0.831
                  kc10        230         62      0.999      0.984       0.99      0.847
               command        230         40       0.97          1      0.995      0.811
                   f15        230        123      0.891          1      0.992      0.701
                 kc135        230         91      0.971      0.989      0.986      0.712
                   a10        230         27          1      0.555      0.899      0.456
                    b1        230         20      0.972          1      0.995      0.793
                   aew        230         25      0.945          1       0.99      0.784
                   f22        230         17      0.913          1      0.995      0.725
                    p3        230        105       0.99          1      0.995      0.801
                    p8        230          1      0.637          1      0.995      0.597
                   f35        230         32      0.939      0.938      0.978      0.574
                   f18        230        125      0.985      0.992      0.987      0.817
                   v22        230         41      0.983          1      0.995       0.69
                 su-27        230         31      0.925          1      0.995      0.859
                 il-38        230         27      0.972          1      0.995      0.811
                tu-134        230          1      0.663          1      0.995      0.895
                 su-33        230          2          1      0.611      0.995      0.796
                 an-70        230          2      0.766          1      0.995       0.73
                 tu-22        230         98      0.984          1      0.995      0.831
Speed: 0.2ms preprocess, 3.8ms inference, 0.0ms loss, 0.8ms postprocess per image

改进一:使用UniRepLKNetBlock替换C2f中的Bottleneck,重构C2f模块

测试结果

代码语言:javascript代码运行次数:0运行复制
  Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 15/15 [00:03<00:00,  3.75it/s]
                   all        230       1412      0.966      0.973      0.992      0.769
                   c17        230        131      0.973      0.992      0.995       0.85
                    c5        230         68       0.97      0.968      0.993      0.848
            helicopter        230         43      0.932          1       0.98      0.634
                  c130        230         85      0.996          1      0.995      0.667
                   f16        230         57      0.984      0.965      0.986       0.71
                    b2        230          2      0.897          1      0.995      0.912
                 other        230         86       0.98      0.942      0.977      0.534
                   b52        230         70      0.981      0.986       0.99      0.863
                  kc10        230         62      0.994      0.984      0.989      0.848
               command        230         40       0.99          1      0.995      0.845
                   f15        230        123      0.989      0.992      0.995      0.688
                 kc135        230         91      0.986      0.989      0.991      0.703
                   a10        230         27          1      0.715      0.969       0.49
                    b1        230         20      0.985          1      0.995      0.745
                   aew        230         25      0.938          1      0.995      0.786
                   f22        230         17      0.925          1      0.995      0.795
                    p3        230        105          1      0.995      0.995      0.788
                    p8        230          1      0.809          1      0.995      0.796
                   f35        230         32      0.968      0.953      0.989      0.568
                   f18        230        125      0.988      0.992      0.991      0.844
                   v22        230         41      0.999          1      0.995      0.691
                 su-27        230         31      0.975          1      0.995      0.863
                 il-38        230         27      0.985          1      0.995      0.881
                tu-134        230          1          1          1      0.995      0.895
                 su-33        230          2          1      0.807      0.995      0.759
                 an-70        230          2      0.856          1      0.995      0.895
                 tu-22        230         98      0.995          1      0.995      0.855
Speed: 0.2ms preprocess, 10.8ms inference, 0.0ms loss, 0.7ms postprocess per image

改进二: 使用UniRepLKNetBlock替换主干网络中的C2f模块

测试结果

代码语言:javascript代码运行次数:0运行复制
YOLOv8l summary: 698 layers, 54175376 parameters, 0 gradients, 215.5 GFLOPs
                 Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 15/15 [00:04<00:00,  3.16it/s]
                   all        230       1412      0.971      0.977      0.991      0.752
                   c17        230        131      0.982      0.992      0.995      0.826
                    c5        230         68      0.975      0.985      0.994      0.836
            helicopter        230         43      0.953      0.941      0.968      0.594
                  c130        230         85      0.988      0.977      0.995      0.667
                   f16        230         57      0.979      0.965       0.98      0.677
                    b2        230          2      0.904          1      0.995       0.85
                 other        230         86      0.977      0.942      0.977       0.54
                   b52        230         70      0.994      0.971      0.988      0.832
                  kc10        230         62          1       0.98      0.989      0.836
               command        230         40      0.974          1      0.995      0.828
                   f15        230        123      0.992      0.979      0.995      0.682
                 kc135        230         91      0.993      0.967      0.987       0.69
                   a10        230         27          1      0.753      0.972      0.506
                    b1        230         20          1      0.979      0.995      0.758
                   aew        230         25      0.954          1      0.992      0.776
                   f22        230         17      0.987          1      0.995      0.779
                    p3        230        105          1      0.974      0.995      0.807
                    p8        230          1       0.87          1      0.995      0.697
                   f35        230         32          1       0.98      0.995      0.571
                   f18        230        125      0.991      0.992      0.988      0.827
                   v22        230         41      0.996          1      0.995      0.679
                 su-27        230         31      0.992          1      0.995      0.882
                 il-38        230         27      0.989          1      0.995      0.885
                tu-134        230          1      0.856          1      0.995      0.895
                 su-33        230          2      0.969          1      0.995      0.697
                 an-70        230          2      0.907          1      0.995      0.849
                 tu-22        230         98      0.998          1      0.995      0.831
Speed: 0.2ms preprocess, 14.4ms inference, 0.0ms loss, 0.7ms postprocess per image

和改进一相比,涨点有点少!

总结

本文参与 腾讯云自媒体同步曝光计划,分享自微信公众号。原始发表:2024-02-05,如有侵权请联系 cloudcommunity@tencent 删除测试网络重构

本文标签: YoloV8改进策略UniRepLKNet,大核卷积的最新成果,轻量高效的首选(全网首发)