深度学习AI构图智能裁图显著性检测美感质量评价

Posted 2021-05-25 踟蹰横渡口,彳亍上滩舟。

深度学习、AI构图、智能裁图、显著性检测、美感质量评价

基于美感数据集和改进的Alexnet-SPP的AI构图智能裁图

基于美感数据集和改进的Alexnet-SPP的显著性检测

部分代码下载地址：下载地址

实现思路和流程图

1. 基于显著性数据集，利用Alexnet-SPP回归出图像的显著性区域；
2. 基于显著性区域生成4x4或者5x5的候选框；
3. 利用美感质量数据集训练度量学习Siamese Network，对显著性候选框进行打分。
4. 对打分后的候选框进行排序，进行智能裁图。

效果展示：








测试代码：

import tensorflow as tf
import numpy as np
import cv2
import itertools

tf.reset_default_graph()

def gen_datalabel(filePath):
    f = open(filePath)
    lines = f.readlines()
    lenNum = len(lines)
    Num = 0
    imPath =[]
    imgs = []
    for line in lines:
        Num = Num + 1
        result = line.split(',')
        im = result[0]
        print im
        imPath.append(im)
        print (Num,'--',lenNum, im)
        img = cv2.imread(im)
        img = cv2.resize(img, (640, 480))
        imgs.append(img)
    return imPath, np.asarray(imgs, np.float32)

def generate_box(x, y, z, w, img, imgPath):
    box_centerx = int((x + z) / 2.0)
    box_centery = int((y + w) / 2.0)
    ratiosx = [1.0, 1.05, 1.1, 1.2, 1.3, 1.5]
    ratiosy = [1.0, 1.05, 1.1, 1.2, 1.3, 1.5]
    # ratiosx = [0.9, 0.95, 1.0, 1.05, 1.1, 1.15, 1.2, 1.3, 1.5]
    # ratiosy = [0.9, 0.95, 1.0, 1.05, 1.1, 1.15, 1.2, 1.3, 1.5]
    bbox_result = []
    bbox_result.append(imgPath)
    print (imgPath)
    for ratiox, ratioy in itertools.product(ratiosx, ratiosy):
        boxw = int((z - x) * ratiox)
        boxh = int((w - y) * ratioy)
        gen_boxx = max(box_centerx - (boxw / 2), 0)
        gen_boxy = max(box_centery - (boxh / 2), 0)
        gen_boxz = min(box_centerx + (boxw / 2), img2.shape[1])
        gen_boxw = min(box_centery + (boxh / 2), img2.shape[0])
        bbox_result.append(gen_boxx)
        bbox_result.append(gen_boxy)
        bbox_result.append(gen_boxz)
        bbox_result.append(gen_boxw)
        print(gen_boxx, gen_boxy, gen_boxz, gen_boxw)
        cv2.rectangle(img, (gen_boxx, gen_boxy), (gen_boxz, gen_boxw), (0, 255, 0), 5)
    return img, bbox_result

filePath = 'data_lab.txt'
imPath, data = gen_datalabel(filePath)
bbox_file = open('bbox_file_lab.txt', 'a')
W_image = 640
H_image = 480

with tf.Session() as sess:

    saver = tf.train.import_meta_graph('./model/model.ckpt.meta')
    saver.restore(sess, tf.train.latest_checkpoint('./model/'))

    graph = tf.get_default_graph()
    x_input = graph.get_tensor_by_name("x_input:0")
    keep_prob = graph.get_tensor_by_name("Placeholder_1:0")#attention to modify
    feed_dict = {x_input: data, keep_prob: 0.5}

    logits = graph.get_tensor_by_name("logits_eval:0")

    cls_result = sess.run(logits, feed_dict)

    for i in range(len(imPath)):
        # result_path1 = './lab_data/' + (imPath[i].split('/'))[2]
        # result_path2 = './lab_crop/' + (imPath[i].split('/'))[2]
        result_path3 = './lab_candidate/' + (imPath[i].split('/'))[2]

        img2 = cv2.imread(imPath[i])
        retxx = float(img2.shape[1]) / W_image
        retyy = float(img2.shape[0]) / H_image
        # cv2.rectangle(img2, ( max(int(retxx * cls_result[i][0]), 0), max(int(retyy * cls_result[i][1]), 0)),
        #               ( min(int(retxx * cls_result[i][2]), img2.shape[1]), min(int(retyy * cls_result[i][3]), img2.shape[0])), (0, 0, 0), 5)
        #cv2.imwrite(result_path2, img2)
        xx, yy, ww, hh = int( max(retxx * cls_result[i][0], 0)), int( max(retyy * cls_result[i][1], 0)), int( min(retxx * cls_result[i][2], img2.shape[1])), int( min(retyy * cls_result[i][3], img2.shape[0]))
        img3, bbox_data= generate_box(xx, yy, ww, hh, img2, imPath[i])
        cv2.rectangle(img3, (xx, yy), (ww, hh),(0, 0, 0), 25)
        cv2.imwrite(result_path3, img3)
        for j in range(len(bbox_data)):
            bbox_file.write(str(bbox_data[j]) + ',')
            if j % 4 == 0:
                bbox_file.write('bbox' + str(j / 4) + ',')
        bbox_file.write('\\n')
    bbox_file.close()