photomosaic search

def choose_match_np(lab, img_db, LAB_db, usage_ary, tolerance=1, usage_penalty=1):
    '''
    Uses a numpy data structure to perform matching instead of very slow calls to SQLite. Because the number of images
    in the Images table is the same as the number of rows in the LabColors table, the association is maintained
    via the array indexes.
    :param lab:
    :param img_db: a Python list. Each row is a  row from the sqlite Images database
    :param LAB_db: a numpy array of the LAB colors associated with an image.
    :param tolerance:
    :param usage_penalty:
    :return:
    '''
    JND = 2.3 # "just noticeable difference"
    tol = (JND * tolerance)**2

    lab         = np.asarray(lab,dtype=np.float32)
    search_for  = lab.flatten()

    num_imgs = len(LAB_db)

    rand_noise = np.random.uniform(low=-1.0,high=1.0, size=(num_imgs,1)) * tol
    use_pen    = (usage_ary * usage_penalty) ** 2

    # distance value does not need square root in this case as we're only looking for minimum
    # value, not actual Euclidean Distance.
    dist = np.sum((LAB_db[:,3:] - search_for)**2,axis=1)[:,None] + rand_noise + use_pen

    # find the index corresponding to the minimum distance. This is far cheaper than a sort. Closer
    # to O(n) than O(n log n)
    min_idx = np.unravel_index(dist.argmin(),dist.shape)[0]

    # calculate dL
    L1 = LAB_db[min_idx][3]
    L2 = LAB_db[min_idx][6]
    L3 = LAB_db[min_idx][9]
    L4 = LAB_db[min_idx][12]
    nL1 =search_for[0]
    nL2 =search_for[3]
    nL3 =search_for[6]
    nL4 =search_for[9]

    dL = ((L1-nL1)+(L2-nL2)+(L3-nL3)+(L4-nL4))/4.0
    match =  {'filename':img_db[min_idx][4], 'dL':dL,'image_id':min_idx}
    # increment the usuage index.
    usage_ary[min_idx] += 1
    return min_idx, match