目標檢測數據預處理--尺寸變換

目錄

1. 比例縮放

2. 使用letterbox

---

目標檢測訓練中，我們的數據集尺寸大部分時侯都是不符合網絡輸入的，需要對尺寸進行修改，下面我介紹兩種常用尺寸變換方法：

1. 比例縮放

這種方法就是簡單的對圖片尺寸進行比例縮放，一般使用cv2.resize()對圖片進行縮放，然後計算長寬縮放比例，再通過比例來縮放標註的目標框尺寸。具體代碼如下：

def read_and_resize_picture(img_path, img_boxes):
    """
    :function:讀取圖片，並縮放圖片尺寸到指定尺寸，同時修改目標框尺寸
    :param img_path: 單張圖片路徑
    :param img_boxes: 對應圖片目標框[[xmin,ymin,xmax,ymax],...,[xmin,ymin,xmax,ymax]]
    :return:圖片(已經讀取啦)，修改的目標框
    """
    img = cv2.imread(img_path,flags=1)
    height, width = img.shape[0:2]
    if height == 416 and width == 416:
        return img, img_boxes
    else:
        h_s = height / 416.
        w_s = width / 416.
        # box = img_boxes
        # 將原圖resize成300,300
        img_resize = cv2.resize(img, dsize=(416, 416), interpolation=cv2.INTER_LINEAR)
        # imgs.append(img_resize)
        img_boxes[:, 0:3:2] = img_boxes[:, 0:3:2] / w_s
        img_boxes[:, 1:4:2] = img_boxes[:, 1:4:2] / h_s
        # boxes.append(box)
        return img_resize, img_boxes

這種方法修改圖片尺寸方便，但是會改變圖片特徵

2. 使用letterbox

先對圖片填充黑邊，使圖片尺寸爲正方形，然後再縮放，這樣不會改變圖片尺寸，代碼如下：

import numpy as np
import cv2
import math
def cv2_letterbox_image(image, expected_size, box):
    """
    function:使用letterbox填充並修改圖片尺寸
    :param image: 原圖
    :param expected_size: 想要修改後的尺寸
    :param box: 原圖目標框
    :return: 修改後的圖片，目標框
    """
    ih, iw = image.shape[0:2]
    ew, eh = expected_size
    scale = min(eh / ih, ew / iw)
    nh = int(ih * scale)
    nw = int(iw * scale)
    for i in range(4):
        box[i] = box[i] * scale
    image = cv2.resize(image, (nw, nh), interpolation=cv2.INTER_CUBIC)
    top = (eh - nh) // 2
    box[1] = box[1] + top
    box[3] = box[3] + top
    bottom = eh - nh - top
    left = (ew - nw) // 2
    box[0] = box[0] + left
    box[2] = box[2] + left
    right = ew - nw - left
    new_img = cv2.copyMakeBorder(image, top, bottom, left, right, cv2.BORDER_CONSTANT)
    return new_img, box
def cv2_deletterbox_image(image, current_size, box):
    """
    function:去除圖像灰邊，並還原目標框尺寸
    :param image: 原圖
    :param current_size: 當前圖片尺寸(網絡輸入尺寸)
    :param box: 網絡輸入尺寸下的目標框
    :return: 匹配原圖的目標框
    """
    ih, iw = image.shape[0:2]
    ew, eh = current_size
    scale = min(eh / ih, ew / iw)
    nh = int(ih * scale)
    nw = int(iw * scale)

    top = (eh - nh) // 2
    box[1] = box[1] - top
    box[3] = box[3] - top
    left = (ew - nw) // 2
    box[0] = box[0] - left
    box[2] = box[2] - left
    for i in range(4):
        box[i] = box[i] / scale

    return box
def view_1():
    box = np.array([147, 336, 402, 437])

    ptr1 = (box[0], box[1])
    ptr2 = (box[2], box[3])
    color = (0, 0, 255)
    path = "YOLOV3_high/data/val/JPEG/image_1.jpg"
    img = cv2.imread(path)
    dst_img, box_new = cv2_letterbox_image(img, (415, 415), box)
    cv2.rectangle(img, ptr1, ptr2, color, 2)
    cv2.imshow('src_img', img)
    cv2.rectangle(dst_img, (box_new[0], box_new[1]), (box_new[2], box_new[3]), color, 2)
    cv2.imshow("dst_img", dst_img)
    # cv2.imwrite("pred_img.jpg",dst_img)
    cv2.waitKey(0)
def view_2():
    box = np.array([147, 336, 402, 437])

    ptr1 = (box[0], box[1])
    ptr2 = (box[2], box[3])
    color = (0, 0, 255)
    path = "YOLOV3_high/data/val/JPEG/image_1.jpg"
    img = cv2.imread(path)
    img_src = img.copy()
    dst_img, box_new = cv2_letterbox_image(img, (416, 416), box)
    cv2.rectangle(img, ptr1, ptr2, color, 2)
    cv2.imshow('src_img', img)
    cv2.rectangle(dst_img, (box_new[0], box_new[1]), (box_new[2], box_new[3]), color, 2)
    cv2.imshow("dst_img", dst_img)

    box_src = cv2_deletterbox_image(img_src, (416,416), box_new)
    cv2.rectangle(img_src, (box_src[0], box_src[1]), (box_src[2], box_src[3]), color, 2)
    cv2.imshow("img_src", img_src)
    cv2.waitKey(0)
if __name__ == "__main__":
    # view_1()
    view_2()