利用caffe-ssd对钢材表面缺陷数据集(NEUDataset)进行finetune训练和测试

---

本篇博客主要讲解如何使用在VOC0712数据集下训练好的ssd.caffemodel来对自己的数据集进行finetune训练，涉及到NEUDataset、一些python脚本代码、网络结构的修改、网络训练和测试流程以及批量生成测试图片的python代码。

同时，本篇博客不会再涉及到路径和名称修改问题，关于caffe-ssd的安装编译、网络训练和测试以及路径和名称修改等问题请参考我这篇博客： 目标检测SSD网络在Caffe下的实现

---

数据集来源： NEU surface defect database（
模型：ssd
系统平台：linux-ubuntu

---

NEUDataset介绍

该数据集是东北大学宋克臣团队制作而成，是钢材表面缺陷数据集，共有1800张图片，包含六种类型：

• crazing
• inclusion
• patches
• pitted_surface
• rolled-in_scale
• scratches

LMDB数据集制作

将数据集分为trainval和test

我写了个python脚本，将1800张images和labels按照8:2的比例随机分为trainval和test两个数据集

• trainval:1440
• test:360
  代码如下：

"""this code is to split randomly images and xml files to train and test file"""

    import os
    import cv2
    #import string
    import random
    import numpy as np
    import shutil
    
    os.makedirs('/home1/xxx/caffe_ssd/data/NEU/neu/trainval/ANNOTATIONS')
    os.makedirs('/home1/xxx/caffe_ssd/data/NEU/neu/trainval/IMAGES')
    os.makedirs('/home1/xxx/caffe_ssd/data/NEU/neu/test/ANNOTATIONS')
    os.makedirs('/home1/xxx/caffe_ssd/data/NEU/neu/test/IMAGES')
    
    open_dir = "/home1/xxx/caffe_ssd/data/NEU/NEU-DET"                           #the file you want to split
    save_dir = '/home1/xxx/caffe_ssd/data/NEU/neu'            #the file you want to save
    sum_samples = 300                                                 #the sums of each class
    img_resize = 300
    sample_class=['crazing', 'inclusion', 'patches', 'pitted_surface', 'rolled-in_scale', 'scratches']   #samples class
    
    def get_specific_suffix(dirname, suffix='.jpg'):     #get specific suffix images and xml files
        images_path = 'IMAGES'                           #the file name of images
        annotations_path = 'ANNOTATIONS'                 #the file name of annotations
    img_dir = os.path.join(dirname, images_path)
    img_list = os.listdir(img_dir)

    xml_dir = os.path.join(dirname, annotations_path)
    xml_list = os.listdir(xml_dir)

    img_list_suffix = []
    for img_array in img_list:
        if os.path.splitext(img_array)[1] == suffix:
            img_list_suffix.append(img_array)
        else:
            continue

    return img_list_suffix, xml_list  #['crazing_1.jpg', 'crazing_10.jpg']   #return img list and xml list of content

    def get_random_list(sum_samples, scale=0.8):     #get random list to split train and test with scale
    list_random = random.sample(range(1, sum_samples), int(sum_samples * scale))   #get random figures without repetition
    list_sort = sorted(list_random)
    return list_sort

    #get random images and annotations, split them to train and test file

    def get_random_img_anno(img_list_suffix, xml_list, sum_samples, img_anno_path='./data/', save_path='./'):
        images_path = 'IMAGES/'              #the file name of images
        annotations_path = 'ANNOTATIONS/'    #the file name of annotations

    random_list = get_random_list(sum_samples)  #get random list
    #split images to train and test according sample class
    for sam_class in sample_class:
        
        for img_name in img_list_suffix:
            count = 0
            
            for i in random_list:           
                if img_name.find(sam_class) != -1:
                    
                    if img_name.split('.')[0] == sam_class + '_' + str(i):
                        shutil.copy(os.path.join(img_anno_path, images_path, img_name),os.path.join(save_path, 'trainval/', images_path, img_name))
                    if img_name.split('.')[0] != sam_class + '_' + str(i):
                        count = count + 1
                        if count == len(random_list):
                            count = 0
                            shutil.copy(os.path.join(img_anno_path, images_path, img_name),os.path.join(save_path, 'test/', images_path, img_name))

    #split annotations to train and test according sample class
    for sam_class in sample_class:
        #count_val = 0
        for xml_name in xml_list:
            count = 0
            
            for i in random_list:
                if xml_name.find(sam_class)