Algorithm 2 The pseudocode of our proposed CME-CNN

Input:

$Dection model index u and e;$

$NMS threshold N_t, Score threshold s;$

Output:

$The detection result imgBuffer;$

1: /*$Bulid model and load model parameter$*/

2: $segamentation Model = nets Factory.getModel\left(u\right);$

3. $detection Model = nets Factory.getModel\left(e\right);$

4: $imgBuffer \leftarrow \left\{ \right\};$

5: /*$Detection Framework Start$*/

6: while $detection Model On$ do

7:    /* $Get Img Frame*/$

8:    $img=camera.getImgFrame();$

9:    $/*Preprocess high resolution image*/$

10:    $imgBlocks=Preprocess\left(img\right) imgBlock\in R^{4\times 256\times 256\times 3}$

11.    $/*Morphological filter processing*/$

12:    $refineBlocks = Filter\left(maskBlocks\right), refineBlocks\in R^{4\times 256\times 256\times 3}$

13:    $/*Get the img without background */$

14:    $maskImg = Combine\left(refineBlocks.*imgBlocks\right), maskImg\in R^{1024\times 1024\times 3}$

15.    $output=detection Model.detect\left(maskImg\right),$         $output\in R^{200\times 6}$

16.    /* $Decode the output$*/

17.    $Boxes,Score=Decode output\left(output\right)$

18.    /*$Non Maximum Suppression$*/

19.    $Boxes,Score=NMS\left(Boxes,Score,N_t\right)$

20.    for $each \left(box,score\right)\in \left(Boxes,Score\right)$ do

21.        /*$Draw the Bounding Boxes in$*/

22:        if $score\ge s$ then

23:         $resultImg=Rectangle\left(reshapeImg,box\right)$

24:        /*Append the result in img buffer*/

25:           $Imgbuffer\leftarrow imgBuffer\cup resultImg$

26:        end if

27:    end for

28:    if $detection Mode off$ then

29:        break;

30:    end if

31: end while