Fig. 1. Framework of our algorithm
Fig. 2. Schematic diagram of convolution decomposition. (a) Standard convolution process; (b) convolution process after decomposition
Fig. 3. Convolutional neural network residualmodule structure diagram
Fig. 4. Deconvolution cascaded structure
Fig. 5. Adaptive candidate region generation
Fig. 6. Visualization detection results of the proposed algorithm in different situations. (a) Small target detection results; (b) dense target detection results;(c) detection results of target under different illuminations
Layer | Type | Kernel | Output size | Number of output channels |
---|
X | input | | 224×224 | 3 | Conv_1 | Convolution | 3×3,64 stride 2 | 112×112 | 32 | Conv_2 | Convolution | ×3 | 56×56 | 64 | Conv_3 | Convolution | ×4 | 28×28 | 128 | Conv_4 | Convolution | ×6 | 14×14 | 256 | Conv_5 | Convolution | ×3 | 7×7 | 512 |
|
Table 1. Lightweight deep residual network model
Layer | Type | Kernel | Stride | Output size |
---|
h1 | Deconvolution | 3×3 | 1 | 14×14×256 | h2 | Deconvolution | 3×3 | 1 | 28×28×256 | h3 | Deconvolution | 3×3 | 1 | 56×56×256 |
|
Table 2. Deconvolution layer parameters
Model | Size /MB | Ratio /% | Accuracy /% |
---|
Resnet | 97.7 | — | 81.3 | LResnet | 10.2 | 10.4 | 80.6 |
|
Table 3. Feature extraction network comparison
Method | mAP | AP50 | AP75 |
---|
①Faster-RCN(Resnet50+RPN) | 18.63 | 35.87 | 17.86 | ②LResnet+RPN | 18.52 | 35.75 | 17.44 | ③LResnet+DC+RPN | 21.03 | 38.46 | 18.03 | ④LResnet+DC+GA-RPN(ours) | 22.12 | 38.76 | 21.53 |
|
Table 4. Effectiveness test of each module for different methods%
Method | Pedestrian | Person | Bicycle | Car | Van | Truck | Tricycle | Awn | Bus | Motor |
---|
Faster-RCNN | 18.34 | 7.62 | 6.76 | 43.31 | 27.53 | 19.95 | 10.13 | 7.65 | 36.87 | 8.79 | Ours | 22.43 | 7.61 | 8.56 | 50.18 | 34.63 | 24.34 | 14.11 | 9.08 | 36.25 | 14.88 |
|
Table 5. Comparison between the results of ten categories from ours model and Faster-RCNN on VisDrone dataset%
Method | mAP /% | AP50 /% | AP75 /% | Frame rate /(frame·s-1) |
---|
FPN | 16.51 | 32.20 | 14.91 | 6 | YOLOv3 | 20.30 | 44.12 | 15.80 | 44 | RetinaNet | 11.81 | 21.37 | 11.62 | 11 | CornerNet | 17.41 | 34.12 | 15.78 | 13 | Ours | 22.12 | 38.76 | 21.53 | 24 |
|
Table 6. Comparison test of UAV aerial data with mainstream object detection algorithm