关键词：Retinex 、 数据扩增 、 颜色恢复 、 对比度增强

前言

  随着深度学习技术的发展，数据扩增已经成为了训练高质量模型的重要步骤之一。然而，传统的数据扩增方法往往不能完全满足需求，因此，研究人员开始探索更加高效和有效的数据扩增方法。本文将介绍一种基于retinex算法的数据扩增方法，该方法可以在不影响图像内容的前提下，提高图像的亮度、对比度和颜色饱和度等方面的表现，从而提高模型的训练效果。

retinex

  基于视网膜理论公式：图像=反射率 * 亮度，我们需要的是估计亮度函数。该技术适用于高动态范围图像增强、水下图像增强、雾天图像增强和弱光图像增强。

在这里我们以下图作为原图进行图像变换：

retinex_FM实现

def retinex_FM(img,iter=4):
    if len(img.shape)==2:
        img=img[...,None]
    ret=np.zeros(img.shape,dtype='uint8')
    def update_OP(x,y):
        nonlocal OP
        IP=OP.copy()
        if x>0 and y==0:
            IP[:-x,:]=OP[x:,:]+R[:-x,:]-R[x:,:]
        if x==0 and y>0:
            IP[:,y:]=OP[:,:-y]+R[:,y:]-R[:,:-y]
        if x<0 and y==0:
            IP[-x:,:]=OP[:x,:]+R[-x:,:]-R[:x,:]
        if x==0 and y<0:
            IP[:,:y]=OP[:,-y:]+R[:,:y]-R[:,-y:]
        IP[IP>maximum]=maximum
        OP=(OP+IP)/2
    for i in range(img.shape[-1]):
        R=np.log(img[...,i].astype('double')+1)
        maximum=np.max(R)
        OP=maximum*np.ones(R.shape)
        S=2**(int(np.log2(np.min(R.shape))-1))
        while abs(S)>=1: #iterations is slow
            for k in range(iter):
                update_OP(S,0)
                update_OP(0,S)
            S=int(-S/2)
        OP=np.exp(OP)
        mmin=np.min(OP)
        mmax=np.max(OP)
        ret[...,i]=(OP-mmin)/(mmax-mmin)*255
    return ret.squeeze()

retinex_SSR实现

def retinex_SSR(img,sigma):
    
    if len(img.shape)==2:
        img=img[...,None]
    ret=np.zeros(img.shape,dtype='uint8')
    for i in range(img.shape[-1]):
        channel=img[...,i].astype('double')
        S_log=np.log(channel+1)
        gaussian=gauss_blur(channel,sigma)
        #gaussian=cv2.filter2D(channel,-1,get_gauss_kernel(sigma)) #conv may be slow if size too big
        #gaussian=cv2.GaussianBlur(channel,(0,0),sigma) #always slower
        L_log=np.log(gaussian+1)
        r=S_log-L_log
        R=r #R=np.exp(r)?
        mmin=np.min(R)
        mmax=np.max(R)
        stretch=(R-mmin)/(mmax-mmin)*255 #linear stretch
        ret[...,i]=stretch
    return ret.squeeze()

retinex_MSR实现

def retinex_MSR(img,sigmas=[15,80,250],weights=None):
    '''r=∑(log(S)-log(S*G))w, MSR combines various SSR with different(or same) weights, 
       commonly we select 3 scales(sigma) and equal weights, (15,80,250) is a good 
       choice. If len(sigmas)=1, equal to SSR
    args:
       sigmas: a list
       weights: None or a list, it represents the weight for each SSR, their sum should 
          be 1, if None, the weights will be [1/t, 1/t, ..., 1/t], t=len(sigmas)
    '''
    if weights==None:
        weights=np.ones(len(sigmas))/len(sigmas)
    elif not abs(sum(weights)-1)<0.00001:
        raise ValueError('sum of weights must be 1!')
    ret=np.zeros(img.shape,dtype='uint8')
    if len(img.shape)==2:
        img=img[...,None]
    for i in range(img.shape[-1]):
        channel=img[...,i].astype('double')
        r=np.zeros_like(channel)
        for k,sigma in enumerate(sigmas):
            r+=(np.log(channel+1)-np.log(gauss_blur(channel,sigma,)+1))*weights[k]
        mmin=np.min(r)
        mmax=np.max(r)
        stretch=(r-mmin)/(mmax-mmin)*255
        ret[...,i]=stretch
    return ret.squeeze()

retinex_gimp实现

def retinex_gimp(img,sigmas=[12,80,250],dynamic=2):
    alpha=128
    gain=1
    offset=0
    img=img.astype('double')+1 #
    csum_log=np.log(np.sum(img,axis=2))
    msr=MultiScaleRetinex(img-1,sigmas) #-1
    r=gain*(np.log(alpha*img)-csum_log[...,None])*msr+offset
    mean=np.mean(r,axis=(0,1),keepdims=True)
    var=np.sqrt(np.sum((r-mean)**2,axis=(0,1),keepdims=True)/r[...,0].size)
    mmin=mean-dynamic*var
    mmax=mean+dynamic*var
    stretch=(r-mmin)/(mmax-mmin)*255
    stretch[stretch>255]=255
    stretch[stretch<0]=0
    return stretch.astype('uint8')

retinex_MSRCR实现

def retinex_MSRCR(img,sigmas=[12,80,250],s1=0.01,s2=0.01):
    alpha=125
    img=img.astype('double')+1 #
    csum_log=np.log(np.sum(img,axis=2))
    msr=MultiScaleRetinex(img-1,sigmas) #-1
    r=(np.log(alpha*img)-csum_log[...,None])*msr
    #beta=46;G=192;b=-30;r=G*(beta*r-b) #deprecated
    #mmin,mmax=np.min(r),np.max(r)
    #stretch=(r-mmin)/(mmax-mmin)*255 #linear stretch is unsatisfactory
    for i in range(r.shape[-1]):
        r[...,i]=simplest_color_balance(r[...,i],0.01,0.01)
    return r.astype('uint8')

retinex_MSRCP算法

def retinex_MSRCP(img,sigmas=[12,80,250],s1=0.01,s2=0.01):
    Int=np.sum(img,axis=2)/3
    Diffs=[]
    for sigma in sigmas:
        Diffs.append(np.log(Int+1)-np.log(gauss_blur(Int,sigma)+1))
    MSR=sum(Diffs)/3
    Int1=simplest_color_balance(MSR,s1,s2)
    B=np.max(img,axis=2)
    A=np.min(np.stack((255/(B+eps),Int1/(Int+eps)),axis=2),axis=-1)
    return (A[...,None]*img).astype('uint8')

cv2_heq实现

def cv2_heq(img,yuv=False):
    if len(img.shape)==2:
        img=img[...,None]
    if yuv:
        img=cv2.cvtColor(img,cv2.COLOR_BGR2YCrCb)
    ret=img.copy()
    for i in range(img.shape[-1]):
        ret[...,i]=cv2.equalizeHist(img[...,i])
        if yuv:
            break
    if yuv:
        return cv2.cvtColor(ret,cv2.COLOR_YCrCb2BGR)
    return ret.squeeze()

retinex_AMSR算法

def retinex_AMSR(img,sigmas=[12,80,250]):
    img=img.astype('double')+1 #
    msr=MultiScaleRetinex(img-1,sigmas,flag=False) #
    y=0.05
    for i in range(msr.shape[-1]):
        v,c=np.unique((msr[...,i]*100).astype('int'),return_counts=True)
        sort_v_index=np.argsort(v)
        sort_v,sort_c=v[sort_v_index],c[sort_v_index] #plot hist
        zero_ind=np.where(sort_v==0)[0][0]
        zero_c=sort_c[zero_ind]
        #
        _=np.where(sort_c[:zero_ind]<=zero_c*y)[0]
        if len(_)==0:
            low_ind=0
        else:
            low_ind=_[-1]
        _=np.where(sort_c[zero_ind+1:]<=zero_c*y)[0]
        if len(_)==0:
            up_ind=len(sort_c)-1
        else:
            up_ind=_[0]+zero_ind+1
        #
        low_v,up_v=sort_v[[low_ind,up_ind]]/100 #low clip value and up clip value
        msr[...,i]=np.maximum(np.minimum(msr[:,:,i],up_v),low_v)
        mmin=np.min(msr[...,i])
        mmax=np.max(msr[...,i])
        msr[...,i]=(msr[...,i]-mmin)/(mmax-mmin)*255
    msr=msr.astype('uint8')
    return msr

结尾

  Retinex算法在数据扩增领域的应用是十分广泛的。通过对图像的处理，我们可以得到更加清晰、明亮、自然的图像，从而提升了机器学习模型的准确度和稳定性。而且，Retinex算法的优点在于可以针对不同的图像进行不同的处理，满足了数据扩增的个性化需求。因此，在进行图像数据扩增时，Retinex算法是一种十分有效的方法，值得我们深入研究和应用。