新公司怎么做网站,雄县网站制作建设中心,wordpress前端发送后端,泉州晋江网站建设费用Selective Search 背景:事先不知道需要检测哪个类别,且候选目标存在层级关系与尺度关系 常规解决方法#xff1a;穷举法#xff0c;在原始图片上进行不同尺度不同大小的滑窗#xff0c;获取每个可能的位置 弊端#xff1a;计算量大#xff0c;且尺度不能兼顾 Selective …Selective Search 背景:事先不知道需要检测哪个类别,且候选目标存在层级关系与尺度关系 常规解决方法穷举法·在原始图片上进行不同尺度不同大小的滑窗获取每个可能的位置 弊端计算量大且尺度不能兼顾 Selective Search:通过视觉特征减少分类可能性 算法步骤 基于图的图像分割方法初始化区域图像分割为很多很多小块循环 使用贪心策略计算相邻区域相似度每次合并相似的两块直到剩下一块 结束 如何保证特征多样性 颜色空间变换RGB,i,Lab,HSV, 距离计算方式 颜色距离 计算每个通道直方图取每个对应bins的直方图最小值直方图大小加权区域/总区域 纹理距离 计算每个区域的快速sift特征(方向个数为8)每个通道bins为2其他用颜色距离 优先合并小区域 单纯通过颜色和纹理合并 合并区域会不断吞并造成多尺度应用在局部问题上无法全局多尺度解决方法给小区域更多权重 .区域的合适度度距离 除了考虑每个区域特征的吻合程度还要考虑区域吻合度(合并后的区域尽量规范不能出现断崖式的区域)直接需求就是区域的外接矩形的重合面积要大 加权综合衡量距离 给予各种距离整合一些区域建议加权综合考虑 参数初始化多样性 通过多种参数初始化图像分割 区域打分 代码实现
# -*- coding: utf-8 -*-
from __future__ import divisionimport cv2 as cv
import skimage.io
import skimage.feature
import skimage.color
import skimage.transform
import skimage.util
import skimage.segmentation
import numpy# Selective Search for Object Recognition by J.R.R. Uijlings et al.
#
# - Modified version with LBP extractor for texture vectorizationdef _generate_segments(im_orig, scale, sigma, min_size):segment smallest regions by the algorithm of Felzenswalb andHuttenlocher# open the Imageim_mask skimage.segmentation.felzenszwalb(skimage.util.img_as_float(im_orig), scalescale, sigmasigma,min_sizemin_size)# merge mask channel to the image as a 4th channelim_orig numpy.append(im_orig, numpy.zeros(im_orig.shape[:2])[:, :, numpy.newaxis], axis2)im_orig[:, :, 3] im_maskreturn im_origdef _sim_colour(r1, r2):calculate the sum of histogram intersection of colourreturn sum([min(a, b) for a, b in zip(r1[hist_c], r2[hist_c])])def _sim_texture(r1, r2):calculate the sum of histogram intersection of texturereturn sum([min(a, b) for a, b in zip(r1[hist_t], r2[hist_t])])def _sim_size(r1, r2, imsize):calculate the size similarity over the imagereturn 1.0 - (r1[size] r2[size]) / imsizedef _sim_fill(r1, r2, imsize):calculate the fill similarity over the imagebbsize ((max(r1[max_x], r2[max_x]) - min(r1[min_x], r2[min_x]))* (max(r1[max_y], r2[max_y]) - min(r1[min_y], r2[min_y])))return 1.0 - (bbsize - r1[size] - r2[size]) / imsizedef _calc_sim(r1, r2, imsize):return (_sim_colour(r1, r2) _sim_texture(r1, r2) _sim_size(r1, r2, imsize) _sim_fill(r1, r2, imsize))def _calc_colour_hist(img):calculate colour histogram for each regionthe size of output histogram will be BINS * COLOUR_CHANNELS(3)number of bins is 25 as same as [uijlings_ijcv2013_draft.pdf]extract HSVBINS 25hist numpy.array([])for colour_channel in (0, 1, 2):# extracting one colour channelc img[:, colour_channel]# calculate histogram for each colour and join to the resulthist numpy.concatenate([hist] [numpy.histogram(c, BINS, (0.0, 255.0))[0]])# L1 normalizehist hist / len(img)return histdef _calc_texture_gradient(img):calculate texture gradient for entire imageThe original SelectiveSearch algorithm proposed Gaussian derivativefor 8 orientations, but we use LBP instead.output will be [height(*)][width(*)]ret numpy.zeros((img.shape[0], img.shape[1], img.shape[2]))for colour_channel in (0, 1, 2):ret[:, :, colour_channel] skimage.feature.local_binary_pattern(img[:, :, colour_channel], 8, 1.0)# LBP特征return retdef _calc_texture_hist(img):calculate texture histogram for each regioncalculate the histogram of gradient for each coloursthe size of output histogram will beBINS * ORIENTATIONS * COLOUR_CHANNELS(3)BINS 10hist numpy.array([])for colour_channel in (0, 1, 2):# mask by the colour channelfd img[:, colour_channel]# calculate histogram for each orientation and concatenate them all# and join to the resulthist numpy.concatenate([hist] [numpy.histogram(fd, BINS, (0.0, 1.0))[0]])# L1 Normalizehist hist / len(img)return histdef _extract_regions(img):R {}# get hsv imagehsv skimage.color.rgb2hsv(img[:, :, :3])# pass 1: count pixel positionsfor y, i in enumerate(img):for x, (r, g, b, l) in enumerate(i):# initialize a new regionif l not in R:R[l] {min_x: 0xffff, min_y: 0xffff,max_x: 0, max_y: 0, labels: [l]}# bounding boxif R[l][min_x] x:R[l][min_x] xif R[l][min_y] y:R[l][min_y] yif R[l][max_x] x:R[l][max_x] xif R[l][max_y] y:R[l][max_y] y# pass 2: calculate texture gradienttex_grad _calc_texture_gradient(img)# pass 3: calculate colour histogram of each regionfor k, v in list(R.items()):# colour histogrammasked_pixels hsv[:, :, :][img[:, :, 3] k]R[k][size] len(masked_pixels / 4)R[k][hist_c] _calc_colour_hist(masked_pixels)# texture histogramR[k][hist_t] _calc_texture_hist(tex_grad[:, :][img[:, :, 3] k])return Rdef _extract_neighbours(regions):def intersect(a, b):if (a[min_x] b[min_x] a[max_x]and a[min_y] b[min_y] a[max_y]) or (a[min_x] b[max_x] a[max_x]and a[min_y] b[max_y] a[max_y]) or (a[min_x] b[min_x] a[max_x]and a[min_y] b[max_y] a[max_y]) or (a[min_x] b[max_x] a[max_x]and a[min_y] b[min_y] a[max_y]):return Truereturn FalseR list(regions.items())neighbours []for cur, a in enumerate(R[:-1]):for b in R[cur 1:]:if intersect(a[1], b[1]):neighbours.append((a, b))return neighboursdef _merge_regions(r1, r2):new_size r1[size] r2[size]rt {min_x: min(r1[min_x], r2[min_x]),min_y: min(r1[min_y], r2[min_y]),max_x: max(r1[max_x], r2[max_x]),max_y: max(r1[max_y], r2[max_y]),size: new_size,hist_c: (r1[hist_c] * r1[size] r2[hist_c] * r2[size]) / new_size,hist_t: (r1[hist_t] * r1[size] r2[hist_t] * r2[size]) / new_size,labels: r1[labels] r2[labels]}return rtdef selective_search(im_orig, scale1.0, sigma0.8, min_size50):Selective SearchParameters----------im_orig : ndarrayInput imagescale : intFree parameter. Higher means larger clusters in felzenszwalb segmentation.sigma : floatWidth of Gaussian kernel for felzenszwalb segmentation.min_size : intMinimum component size for felzenszwalb segmentation.Returns-------img : ndarrayimage with region labelregion label is stored in the 4th value of each pixel [r,g,b,(region)]regions : array of dict[{rect: (left, top, width, height),labels: [...],size: component_size},...]# 期待输入3通道图片assert im_orig.shape[2] 3, 3ch image is expected# load image and get smallest regions# region label is stored in the 4th value of each pixel [r,g,b,(region)]# 基于图方法生成图的最小区域img _generate_segments(im_orig, scale, sigma, min_size)# (512, 512, 4)# print(img.shape)# cv2.imshow(res1, im_orig)# print(type(img))# # img cv2.cvtColor(img,cv2.COLOR_RGB2BGR)# cv2.imshow(res,img)# cv2.waitKey(0)# # print(img)# exit()if img is None:return None, {}imsize img.shape[0] * img.shape[1]# 拓展区域R _extract_regions(img)# extract neighbouring informationneighbours _extract_neighbours(R)# calculate initial similaritiesS {}for (ai, ar), (bi, br) in neighbours:S[(ai, bi)] _calc_sim(ar, br, imsize)# hierarchal searchwhile S ! {}:# get highest similarityi, j sorted(S.items(), keylambda i: i[1])[-1][0]# merge corresponding regionst max(R.keys()) 1.0R[t] _merge_regions(R[i], R[j])# mark similarities for regions to be removedkey_to_delete []for k, v in list(S.items()):if (i in k) or (j in k):key_to_delete.append(k)# remove old similarities of related regionsfor k in key_to_delete:del S[k]# calculate similarity set with the new regionfor k in [a for a in key_to_delete if a ! (i, j)]:n k[1] if k[0] in (i, j) else k[0]S[(t, n)] _calc_sim(R[t], R[n], imsize)regions []for k, r in list(R.items()):regions.append({rect: (r[min_x], r[min_y],r[max_x] - r[min_x], r[max_y] - r[min_y]),size: r[size],labels: r[labels]})return img, regions测试
# -*- coding: utf-8 -*-
from __future__ import (division,print_function,
)
import cv2 as cvimport skimage.data
import matplotlib.pyplot as plt
import matplotlib.patches as mpatches
import selectivesearchdef main():# loading astronaut imageimg skimage.data.astronaut()# print(type(img))# img cv.cvtColor(img,cv.COLOR_RGB2BGR)# cv.imshow(res,img)# cv.waitKey(0)# # print(img)# exit()# perform selective searchimg_lbl, regions selectivesearch.selective_search(img, scale500, sigma0.9, min_size10)candidates set()for r in regions:# excluding same rectangle (with different segments)if r[rect] in candidates:continue# excluding regions smaller than 2000 pixelsif r[size] 2000:continue# distorted rectsx, y, w, h r[rect]if w / h 1.2 or h / w 1.2:continuecandidates.add(r[rect])# draw rectangles on the original imagefig, ax plt.subplots(ncols1, nrows1, figsize(6, 6))ax.imshow(img)for x, y, w, h in candidates:print(x, y, w, h)rect mpatches.Rectangle((x, y), w, h, fillFalse, edgecolorred, linewidth1)ax.add_patch(rect)plt.show()if __name__ __main__:main() 测试结果
RCNN
算法步骤 产生目标区域候选 CNN目标特征提取 使用的AlexNetimageNet预训练迁移学习只训练全连接层采用的全连接层输出(导致输入大小必须固定) 目标种类分类器 SVM困难样本挖掘方法正样本—正样本 iou0.3 负样本 贪婪非极大值抑制 NMS 根据分类器的类别分类概率做排序假设从小到大属于正样本的概率 分别为A、B、C、D、E、F。 从最大概率矩形框F开始分别判断A~E与F的重叠度IOU是否大于某个设定的阈值 假设B、D与F的重叠度超过阈值那么就扔掉B、D并标记第一个矩形框F是我们保留下来的。 从剩下的矩形框A、C、E中选择概率最大的E然后判断E与A、C的重叠度重叠度大于一定的阈值那么就扔掉并标记E是我们保留下来的第二个矩形框。 就这样一直重复找到所有被保留下来的矩形框。 BoundingBox回归 微调回归框 一个区域位置 位置映射真实位置 转换偏移量参数 映射关系式 选用pool5层 最小化w 不使用全连接的输出作为非极大抑制的输入而是训练很多的SVM。 因为CNN需要大量的样本当正样本设置为真实BoundingBox时效果很差而IOU0.5相当于30倍的扩充了样本数量。而我们近将CNN结果作为一个初选然后用困难负样本挖掘的SVM作为第二次筛选就好多了 缺点:时间代价太高了
