该系列文章仅限于某验滑块研究,不会公开具体算法源码,欢迎讨论
本文关联文章:
纵观
入微
芥子
浩瀚
一. 识别滑块缺口
- 使用ddddocr识别
该算法识别准确率为95%左右,测试三轮,每轮测试100次
def generate_distance(slice_url, bg_url):
"""
:param bg_url: 背景图地址
:param slice_url: 滑块图地址
:return: distance
:rtype: Integer
"""
slide = ddddocr.DdddOcr(det=False, ocr=False, show_ad=False)
slice_image = requests.get(slice_url).content
bg_image = requests.get(bg_url).content
result = slide.slide_match(target_bytes, bg_image, simple_target=True)
return result['target'][0]
- 使用cv2识别
该算法识别准确率为95%左右,测试三轮,每轮测试100次
def generate_distance(slice_url, bg_url):
"""
:param bg_url: 背景图地址
:param slice_url: 滑块图地址
:return: distance
:rtype: Integer
"""
slice_image = np.asarray(bytearray(requests.get(slice_url).content), dtype=np.uint8)
slice_image = cv2.imdecode(slice_image, 1)
slice_image = cv2.Canny(slice_image, 255, 255)
bg_image = np.asarray(bytearray(requests.get(bg_url).content), dtype=np.uint8)
bg_image = cv2.imdecode(bg_image, 1)
bg_image = cv2.pyrMeanShiftFiltering(bg_image, 5, 50)
bg_image = cv2.Canny(bg_image, 255, 255)
result = cv2.matchTemplate(bg_image, slice_image, cv2.TM_CCOEFF_NORMED)
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(result)
return max_loc[0]
二. 构造滑块轨迹
- 构造轨迹库
图片长度为300,理论上就300种轨迹,实际上应该是200+种,还要减去滑块图的长度80
手动滑他个几百次,并把距离和轨迹记录下来,识别出距离后直接查对应轨迹 - 算法构造轨迹track
def generate_track(distance):
def __ease_out_expo(step):
return 1 if step == 1 else 1 - pow(2, -10 * step)
tracks = [[random.randint(20, 60), random.randint(10, 40), 0]]
count = 30 + int(distance / 2)
_x, _y = 0, 0
for item in range(count):
x = round(__ease_out_expo(item / count) * distance)
t = random.randint(10, 20)
if x == _x:
continue
tracks.append([x - _x, _y, t])
_x = x
tracks.append([0, 0, random.randint(200, 300)])
times = sum([track[2] for track in tracks])
return tracks, times
三. 结语
本篇文章篇幅不长,主要也没啥好说的,验证码研究多了,识别和轨迹就那几套方法,换汤不换药
函数a(e, t)中的重头戏:c.guid()、_.encrypt()、i.encrypt()、c.arrayToHex()四个函数我们放到浩瀚篇再说吧,不然我这紫极魔瞳四大境界变成三大境界了,哈哈哈
浩瀚