基于九宫格与bbox逻辑回归的细粒度 高密度目标检测

其他 2018-11-05 18:11:10 阅读次数: 0
版权声明：我是南七小僧，微信: to_my_love ，欢迎交流思想碰撞。 https://blog.csdn.net/qq_25439417/article/details/83540696 

# -*- coding: utf-8 -*-
"""
Created on Fri Oct 19 21:13:41 2018

@author: Lenovo
"""

from keras.applications.vgg16 import VGG16
from keras.layers import Dense,Conv2D,MaxPooling2D,Average,Concatenate,Reshape,Flatten,Input,Add,Multiply,Lambda,Dropout,TimeDistributed
from keras.models import Model
from keras import backend as K
from keras.optimizers import SGD
from keras.callbacks import ModelCheckpoint,EarlyStopping
from keras.utils import plot_model
from PIL import Image as Im
import numpy as np
import os
import random
import tensorflow as tf
'''
1 2 3
4 0 6
7 8 9
'''

def sum_lambda(a):
    print(a.shape)
    return K.sum(a, axis=-1)

def bbox_loss(yp,yt):
#    print(a.shape)
    return tf.reduce_mean(tf.square(yp-yt))

def get_model():
    vgg16_dense_list = []
    
    input_list = [Input(shape=(50,50,3)) for i in range(9)]
    input_bbox = Input(shape=(4,))
    vgg16 = VGG16(include_top=False, weights='imagenet')
    
    for inp in input_list:
        vgg16_dense = Dense(200,activation='relu')(Flatten()(vgg16(inp)))
        vgg16_dense_list.append(vgg16_dense)
#        print(vgg16_dense.shape)
    sudoku_concat = Concatenate()(vgg16_dense_list)
#    print(vgg16_dense_list[1].shape)
    sudoku_reshape = Reshape((200,9))(sudoku_concat)
    
    sudoku_weights = Dense(9,activation='softmax')(sudoku_concat)
#    sudoku_weights_T = Reshape((1,9))(sudoku_weights)
#    
#    print(sudoku_weights[0][0].shape)
    sudoku_multi = Multiply()([sudoku_reshape,sudoku_weights])
    sum_lambda_out = Lambda(sum_lambda)(sudoku_multi)
#    print(sum_lambda_out.shape)
#    sudoku_multi  = Dense(100,activation='relu')(sum_lambda_out )
    drop = Dropout(0.5)(sum_lambda_out)
    out = Dense(1,activation='sigmoid',name='fenlei')(drop)
    
    #bbox 4 parameters
#    drop = Dropout(0.5)(sum_lambda_out)
    bbox100dense = Dense(100,activation='relu')(sum_lambda_out)
    bbox4w = Dense(4,activation='relu')(bbox100dense)
    bboxrs_out = Multiply()([input_bbox,bbox4w])
    bboxout = Dense(4,activation='sigmoid',name='huigui')(bboxrs_out)
    
    
    model = Model(inputs=input_list+[input_bbox],outputs=[out,bboxout])
    
    model.compile(loss={'fenlei':'binary_crossentropy',
                        'huigui':'mse'},
                    loss_weights={
                            'fenlei':1.,
                            'huigui':1.
                                  },
                  optimizer=SGD(lr=0.001, momentum=0.99, decay=1e-5),metrics=['accuracy'])
    print(model.summary())
    plot_model(model,to_file='model.png')
    return model

def sudoku_data_generator():
    imglist = get_all_img_path()
    i=0
    while i<len(imglist):
#        x_train = list()
#        y_train = list()
        imgdata = imglist[i].split(' ')
        img = Im.open('data/'+imgdata[0])
        img = np.pad(np.array(img), ((100, 100), (100, 100), (0, 0)), 'constant', constant_values=255)
        img = Im.fromarray(img)
        yuanimg = img
        img = np.array(img)
        x_train0 = []
        x_train1 = []
        x_train2 = []
        x_train3 = []
        x_train4 = []
        x_train6 = []
        x_train7 = []
        x_train8 = []
        x_train9 = []
        x_bbox_train = []
        y_train1 = []
        y_train2 = []
        #append所有人头bbox
        for k in range(2,len(imgdata),5):

            x = int(imgdata[k+1])+100
            y = int(imgdata[k+2])+100
            w = int(imgdata[k+3])
            h = int(imgdata[k+4])
            x_train0.append(np.array(yuanimg.crop((x,y,x+w,y+h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))

            x_train1.append(np.array(yuanimg.crop((x-w,y-h,x,y)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train2.append(np.array(yuanimg.crop((x,y-h,x+w,y)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train3.append(np.array(yuanimg.crop((x+w,y-h,x+2*w,y)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train4.append(np.array(yuanimg.crop((x-w,y,x,y+h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train6.append(np.array(yuanimg.crop((x+w,y,x+2*w,y+h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train7.append(np.array(yuanimg.crop((x-w,y+h,x,y+2*h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train8.append(np.array(yuanimg.crop((x,y+h,x+w,y+2*h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train9.append(np.array(yuanimg.crop((x+w,y+h,x+2*w,y+2*h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_bbox_train.append(np.array([x,y,w,h]))
            
            y_train1.append(np.array([1]))
#            
            y_train2.append(np.array([x/yuanimg.size[0],y/yuanimg.size[1],w/yuanimg.size[0],h/yuanimg.size[1]]))
#            print(np.array([[1]]).shape)
            img[y:y+h,x:x+w]=255
#           _train = list()
#            y_c_train = list()
            
        #随机找五个非人头bbox
        img = Im.fromarray(img)
        for k in range(5):
            x = random.randint(100,img.size[0]-100)
            y = random.randint(100,img.size[1]-100)
            w = random.randint(15,35)
            h = random.randint(15,35)
            x_train0.append(np.array(img.crop((x,y,x+w,y+h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
           
#            12346789
            x_train1.append(np.array(img.crop((x-w,y-h,x,y)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train2.append(np.array(img.crop((x,y-h,x+w,y)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train3.append(np.array(img.crop((x+w,y-h,x+2*w,y)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train4.append(np.array(img.crop((x-w,y,x,y+h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train6.append(np.array(img.crop((x+w,y,x+2*w,y+h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train7.append(np.array(img.crop((x-w,y+h,x,y+2*h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train8.append(np.array(img.crop((x,y+h,x+w,y+2*h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train9.append(np.array(img.crop((x+w,y+h,x+2*w,y+2*h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_bbox_train.append(np.array([x,y,w,h]))
            y_train1.append(np.array([0]))
            y_train2.append(np.array([x/yuanimg.size[0],y/yuanimg.size[1],w/yuanimg.size[0],h/yuanimg.size[1]]))
            
#            x_train.extend(x_c_train)
#            y_train.extend(y_c_train)
            yield [np.array(x_train0),np.array(x_train1),np.array(x_train2),np.array(x_train3),np.array(x_train4),np.array(x_train6),np.array(x_train7),np.array(x_train8),np.array(x_train9),np.array(x_bbox_train)],[np.array(y_train1),np.array(y_train2)]
        i+=1    
        
        
def sudoku_val_data_generator():
    imglist = get_all_img_path()
    i=len(imglist)-300
    while i<len(imglist):
#        x_train = list()
#        y_train = list()
        imgdata = imglist[i].split(' ')
        img = Im.open('data/'+imgdata[0])
        img = np.pad(np.array(img), ((100, 100), (100, 100), (0, 0)), 'constant', constant_values=255)
        img = Im.fromarray(img)
        yuanimg = img
        img = np.array(img)
        x_train0 = []
        x_train1 = []
        x_train2 = []
        x_train3 = []
        x_train4 = []
        x_train6 = []
        x_train7 = []
        x_train8 = []
        x_train9 = []
        x_bbox_train = []
        y_train1 = []
        y_train2 = []
        #append所有人头bbox
        for k in range(2,len(imgdata),5):

            x = int(imgdata[k+1])+100
            y = int(imgdata[k+2])+100
            w = int(imgdata[k+3])
            h = int(imgdata[k+4])
            x_train0.append(np.array(yuanimg.crop((x,y,x+w,y+h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))

            x_train1.append(np.array(yuanimg.crop((x-w,y-h,x,y)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train2.append(np.array(yuanimg.crop((x,y-h,x+w,y)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train3.append(np.array(yuanimg.crop((x+w,y-h,x+2*w,y)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train4.append(np.array(yuanimg.crop((x-w,y,x,y+h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train6.append(np.array(yuanimg.crop((x+w,y,x+2*w,y+h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train7.append(np.array(yuanimg.crop((x-w,y+h,x,y+2*h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train8.append(np.array(yuanimg.crop((x,y+h,x+w,y+2*h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train9.append(np.array(yuanimg.crop((x+w,y+h,x+2*w,y+2*h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_bbox_train.append(np.array([x,y,w,h]))
            
            y_train1.append(np.array([1]))
#            
            y_train2.append(np.array([x/yuanimg.liangd[0],y/yuanimg.size[1],w/yuanimg.size[0],h/yuanimg.size[1]]))
#            print(np.array([[1]]).shape)
            img[y:y+h,x:x+w]=255
#           _train = list()
#            y_c_train = list()
            
        #随机找五个非人头bbox
        img = Im.fromarray(img)
        for k in range(5):
            x = random.randint(100,img.size[0]-100)
            y = random.randint(100,img.size[1]-100)
            w = random.randint(15,35)
            h = random.randint(15,35)
            x_train0.append(np.array(img.crop((x,y,x+w,y+h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
           
#            12346789
            x_train1.append(np.array(img.crop((x-w,y-h,x,y)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train2.append(np.array(img.crop((x,y-h,x+w,y)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train3.append(np.array(img.crop((x+w,y-h,x+2*w,y)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train4.append(np.array(img.crop((x-w,y,x,y+h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train6.append(np.array(img.crop((x+w,y,x+2*w,y+h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train7.append(np.array(img.crop((x-w,y+h,x,y+2*h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train8.append(np.array(img.crop((x,y+h,x+w,y+2*h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_train9.append(np.array(img.crop((x+w,y+h,x+2*w,y+2*h)).resize((50,50),Im.LANCZOS)).reshape((50,50,3)))
            x_bbox_train.append(np.array([x,y,w,h]))
            y_train1.append(np.array([0]))
            y_train2.append(np.array([x/yuanimg.size[0],y/yuanimg.size[1],w/yuanimg.size[0],h/yuanimg.size[1]]))
            
#            x_train.extend(x_c_train)
#            y_train.extend(y_c_train)
            yield [np.array(x_train0),np.array(x_train1),np.array(x_train2),np.array(x_train3),np.array(x_train4),np.array(x_train6),np.array(x_train7),np.array(x_train8),np.array(x_train9),np.array(x_bbox_train)],[np.array(y_train1),np.array(y_train2)]
        i+=1    
        
   
    
def get_all_img_path():   
    path = os.listdir(r'data/')
    print(path)
    imgnamelist=list()
    i = 0     
    for i in path:
        if 'txt' in i:
            with open('data/'+i) as f:
                imgnamelist.extend(f.readlines())
    return imgnamelist

print(1)

mp = ModelCheckpoint(filepath='rentou_best_predict_score_sudoki.h5',
                     save_best_only=True, monitor='val_acc', verbose=1)
cll = [mp]
model = get_model()
model.fit_generator(sudoku_data_generator(),
                  steps_per_epoch=9500,
                  epochs=2,
                  verbose=1,
                  callbacks=cll,
                  validation_data=sudoku_val_data_generator(),
                  validation_steps=100,
                  class_weight=None,
                  shuffle=True)

猜你喜欢

转载自blog.csdn.net/qq_25439417/article/details/83540696
今日推荐
基于大语言模型的开源知识库问答系统 MaxKB GitHub Star 数量突破 5,000 个！
美国拟限制 AI 大模型出口中国和俄罗斯
苹果将与 OpenAI 达成协议，将 ChatGPT 应用于 iPhone
openKylin 社区生态委员会第六次会议圆满召开
阿里云正式发布通义千问 2.5
Python 3.13 发布首个 Beta：实验性自由线程模式和 JIT、改进交互式解释器
Stack Overflow 拿我的代码去训练 AI 大模型，还封了我的账号
Pop!_OS 的 COSMIC 桌面完成 App Store 上架工作
《2024 年一季度互联网投融资运行情况》研究报告
报告：Django 仍然是 74% 开发者的首选
15 年前上了“FFmpeg 耻辱柱”，今天他还得谢谢咱——腾讯QQPlayer一雪前耻？
TIOBE 5 月榜单：Fortran “复活”进入 Top 10
周排行
记一下去大梅沙的准备（2018-05-26）
Spring 注解 事务
基于HTTP协议的客户端缓存
阿里云rds 备份和还原
[PHP] 几个拖慢 PHP 程序/API 运行速度的点
python 代码风格------------PEP8规则
js控制json生成菜单——自制菜单（一）
将字符串: 'k:1|k1:2|k2:3|k3:4 ' ,处理成 python 字典: {'k':1, 'k1':2, ...}
微信小程序转支付宝小程序
Qt551.窗口滚动条
每日归档
更多
2024-05-13(18)
2024-05-12(0)
2024-05-11(38)
2024-05-10(38)
2024-05-09(35)
2024-05-08(42)
2024-05-07(14)
2024-05-06(40)
2024-05-05(0)
2024-05-04(7)

代码天地 © 2018 codetd.com

境外服务器   最新电视剧2022