PyCS/models/moth_scanner/scanner/detector.py

import cv2
import numpy as np
import typing as T

from munch import munchify
from collections import namedtuple
from skimage import filters

# the coordinates are relative!
BBox = namedtuple("BBox", "x0 y0 x1 y1")
BBoxInfo = namedtuple("BBoxInfo", "area ratio mean std selected", defaults=[-1, -1, -1, -1, False])
Detection = namedtuple("Detection", "bbox info")




class Detector(object):


    def __init__(self, configuration: T.Dict[str, T.Dict]) -> None:
        super().__init__()
        config = munchify(configuration)

        self.scale: float = config.preprocess.scale
        self.min_size: int = config.preprocess.min_size
        self.sigma: float = config.preprocess.sigma

        self.block_size_scale: float = config.threshold.block_size_scale

        self.dilate_iterations: int = config.postprocess.dilate_iterations
        self.kernel_size: int = config.postprocess.kernel_size


    def __call__(self, im: np.ndarray) -> T.List[Detection]:

        _im = self.rescale(im)

        im0 = self.preprocess(_im)
        im1 = self.threshold(im0)
        im2 = self.postprocess(im1)

        bboxes = self.detect(im2)

        return self.postprocess_boxes(_im, bboxes)


    def detect(self, im: np.ndarray) -> T.List[BBox]:

        contours, hierarchy = cv2.findContours(im, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
        contours = sorted(contours, key=cv2.contourArea, reverse=True)

        return [_contour2bbox(c, im.shape) for c in contours]


    def rescale(self, im: np.ndarray) -> np.ndarray:

        H, W = im.shape
        _scale = self.min_size / min(H, W)
        scale = max(self.scale, min(1, _scale))
        size = int(W * scale), int(H * scale)

        return cv2.resize(im, dsize=size)


    def preprocess(self, im: np.ndarray) -> np.ndarray:
        res = filters.gaussian(im, sigma=self.sigma, preserve_range=True)
        return res.astype(im.dtype)


    def threshold(self, im: np.ndarray) -> np.ndarray:
        block_size_scale = self.block_size_scale

        # make block size an odd number
        block_size = min(im.shape) * block_size_scale // 2 * 2 + 1

        thresh = filters.threshold_local(im,
            block_size=block_size,
            mode="constant",
        )

        max_value = 255
        bin_im = ((im > thresh) * max_value).astype(np.uint8)
        return max_value - bin_im


    def postprocess(self, im: np.ndarray) -> np.ndarray:
        kernel_size = self.kernel_size
        iterations = self.dilate_iterations
        kernel = np.ones((kernel_size, kernel_size), dtype=np.uint8)

        im = cv2.morphologyEx(im, cv2.MORPH_OPEN, kernel)
        im = cv2.morphologyEx(im, cv2.MORPH_CLOSE, kernel)

        if iterations >= 1:
            im = cv2.erode(im, kernel, iterations=iterations)
            im = cv2.dilate(im, kernel, iterations=iterations)

        return im


    def postprocess_boxes(self, im: np.ndarray, bboxes: T.List[BBox]):

        detections = [Detection(bbox, BBoxInfo()) for bbox in bboxes]

        _im = im.astype(np.float64) / 255.
        integral, integral_sq = cv2.integral2(_im)
        # im_mean, im_std, im_n = _im_mean_std(integral, integral_sq)

        inds = cv2.dnn.NMSBoxes([[x0, y0, x1-x0, y1-y0] for (x0, y0, x1, y1) in bboxes],
                            np.ones(len(bboxes), dtype=np.float32),
                            score_threshold=0.99,
                            nms_threshold=0.1,
                           )

        # calculate the BBoxInfos only for the selected and update the detections
        for i in inds.squeeze():
            bbox, _ = detections[i]
            mean, std, n = _im_mean_std(integral, integral_sq, bbox)
            area, ratio = _area(bbox), _ratio(bbox)
            selected = self.is_selected(mean, std, ratio, area)
            info = BBoxInfo(mean, std, area, ratio, selected)
            detections[i] = Detection(bbox, info)

        return detections

    def is_selected(self, mean: float, std: float, ratio: float, area: float) -> bool:
        # Caution, here are some magic numbers!
        return \
            std >= 5e-2 and \
            ratio >= 2.5e-1 and \
            4e-4 <= area <= 1/9

def _contour2bbox(contour: np.ndarray, shape: T.Tuple[int, int]) -> BBox:
    """ Gets the maximal extent of a contour and translates it to a bounding box. """
    x0, y0 = contour.min(axis=0)[0].astype(np.int32)
    x1, y1 = contour.max(axis=0)[0].astype(np.int32)

    h, w = shape
    return BBox(x0/w, y0/h, x1/w, y1/h)


def _ratio(bbox: BBox) -> float:
    x0, y0, x1, y1 = bbox
    h, w = y1-y0, x1-x0
    return min(h, w) / max(h, w)


def _area(bbox: BBox) -> float:
    x0, y0, x1, y1 = bbox
    h, w = y1-y0, x1-x0
    return h * w

def _im_mean_std(integral: np.ndarray,
                 integral_sq: np.ndarray,
                 bbox: T.Optional[BBox] = None
                ) -> T.Tuple[float, float, int]:

    h, w = integral.shape[0] - 1, integral.shape[1] - 1

    if bbox is None:
        arr_sum = integral[-1, -1]
        arr_sum_sq = integral_sq[-1, -1]
        N = h * w

    else:
        x0, y0, x1, y1 = bbox
        x0, x1 = int(x0 * w), int(x1 * w)
        y0, y1 = int(y0 * h), int(y1 * h)

        A, B, C, D = (y0,x0), (y1,x0), (y0,x1), (y1,x1)
        arr_sum = integral[D] + integral[A] - integral[B] - integral[C]
        arr_sum_sq = integral_sq[D] + integral_sq[A] - integral_sq[B] - integral_sq[C]

        N = (x1-x0) * (y1-y0)

    arr_mean = arr_sum / N
    arr_std  = np.sqrt((arr_sum_sq - (arr_sum**2) / N) / N)

    return arr_mean, arr_std, N