精读深度学习算法源码系列（一）：keras实现SSD源码:ssd_box_encode_decode_utils

最新推荐文章于 2025-03-11 21:49:30 发布

原创

最新推荐文章于 2025-03-11 21:49:30 发布 · 5.7k 阅读

15 ·

CC 4.0 BY-SA版权

算法原文：SSD: Single Shot MultiBox Detector
github链接：https://github.com/pierluigiferrari/ssd_keras#overview

以ssd7作为主程序一共如下几类：

train_ssd7
keras_ssd7
keras_ssd_loss
ssd_box_encode_decode_utils
ssd_batch_generator

ssd_box_encode_decode_utils

代码主要作用：

对生成的dafult box进行IOU检测
对其进行非极大抑制
将其做成model对SSD进行输出
将其做成model提供给SSD主程序使用

def iou(boxes1, boxes2, coords='centroids'):
# IOU函数:模型产生的目标窗口和原来标记窗口的交叠率
# 矩阵box有两种储存方式：一维状态下包含一个box所需要的坐标数据，2维下是n个box所需要的数据
            if len(boxes1.shape) > 2: raise ValueError("boxes1 must have rank either 1 or 2, but has rank {}.".format(len(boxes1.shape)))
            if len(boxes2.shape) > 2: raise ValueError("boxes2 must have rank either 1 or 2, but has rank {}.".format(len(boxes2.shape)))
# 如果输入了一维矩阵，则对其补0扩维     
            if len(boxes1.shape) == 1: boxes1 = np.expand_dims(boxes1, axis=0)
            if len(boxes2.shape) == 1: boxes2 = np.expand_dims(boxes2, axis=0)
# box是存储有四个方向坐标的，如果缺少坐标则判断错误        
            if not (boxes1.shape[1] == boxes2.shape[1] == 4): raise ValueError("It must be boxes1.shape[1] == boxes2.shape[1] == 4, but it is boxes1.shape[1] == {}, boxes2.shape[1] == {}.".format(boxes1.shape[1], boxes2.shape[1]))
# 判断矩阵的坐标格式：满足则为(cx,cy,w,h),否则(xmin,xmax,ymin,ymax)     
            if coords == 'centroids':

                boxes1 = convert_coordinates(boxes1, start_index=0, conversion='centroids2minmax')
                boxes2 = convert_coordinates(boxes2, start_index=0, conversion='centroids2minmax')
            elif coords != 'minmax':
                raise ValueError("Unexpected value for `coords`. Supported values are 'minmax' and 'centroids'.")
#交集比并集并对其进行非极大抑制，最终返回       
            intersection = np.maximum(0, np.minimum(boxes1[:,1], boxes2[:,1]) - np.maximum(boxes1[:,0], boxes2[:,0])) * np.maximum(0, np.minimum(boxes1[:,3], boxes2[:,3]) - np.maximum(boxes1[:,2], boxes2[:,2]))
            union = (boxes1[:,1] - boxes1[:,0]) * (boxes1[:,3] - boxes1[:,2]) + (boxes2[:,1] - boxes2[:,0]) * (boxes2[:,3] - boxes2[:,2]) - intersection

            return intersection / union

def convert_coordinates(tensor, start_index, conversion='minmax2centroids'):
#转换坐标函数：将box轴对齐，其实就是放在一个中心下，但是转换会有误差
#参数：1.N维包含四个坐标的长量，即N个输入框
#2.在tensor的最后一维进行坐标变换
#3.默认minmax2centroids，对box进行转向
    ind = start_index 
    tensor1 = np.copy(tensor).astype(np.float) #copy后改变不会影响原tensor
    if conversion == 'minmax2centroids':
        tensor1[..., ind] = (tensor[..., ind] + tensor[..., ind+1]) / 2.0 
        tensor1[..., ind+1] = (tensor[..., ind+2] + tensor[..., ind+3]) / 2.0 
        tensor1[..., ind+2] = tensor[..., ind+1] - tensor[..., ind] 
        tensor1[..., ind+3] = tensor[..., ind+3] - tensor[..., ind+2] 
    elif conversion == 'centroids2minmax':
        tensor1[..., ind] = tensor[..., ind] - tensor[..., ind+2] / 2.0 
        tensor1[..., ind+1] = tensor[..., ind] + tensor[..., ind+2] / 2.0  
        tensor1[..., ind+2] = tensor[..., ind+1] - tensor[..., ind+3] / 2.0 
        tensor1[..., ind+3] = tensor[..., ind+1] + tensor[..., ind+3] / 2.0 #
    else:
        raise ValueError("Unexpected conversion value. Supported values are 'minmax2centroids' and 'centroids2minmax'.")

    return tensor1

def convert_coordinates2(tensor, start_index, conversion='minmax2centroids'):
#对上面的convert_coordinates矩阵乘法实现
#这里给出了不适用的原因：偏多的系数矩阵相乘是非常浪费时间并且没意义的
    ind = start_index
    tensor1 = np.copy(tensor).astype(np.float)
    if conversion == 'minmax2centroids':
        M = np.array([[0.5, 0. , -1.,  0.],
                      [0.5, 0. ,  1.,  0.],
                      [0. , 0.5,  0., -1.],
                      [0. , 0.5,  0.,  1.]])
        tensor1[..., ind:ind+4] = np.dot(tensor1[..., ind:ind+4], M)
    elif conversion == 'centroids2minmax':
        M = np.array([[ 1. , 1. ,  0. , 0. ],
                      [ 0. , 0. ,  1. , 1. ],
                      [-0.5, 0.5,  0. , 0. ],
                      [ 0. , 0. , -0.5, 0.5]]) 
        tensor1[..., ind:ind+4] = np.dot(tensor1[..., ind:ind+4], M)
    else:
        raise ValueError("Unexpected conversion value. Supported values are 'minmax2centroids' and 'centroids2minmax'.")

    return tensor1

def greedy_nms(y_pred_decoded, iou_threshold=0.45, coords='minmax'):
'''
NMS（非极大值抑制）：
对输入框向量进行非极大贪婪抑制，主要是消除多余的框，找到最佳匹配的位置
设定一个IOU的阈值，将输入框的A，B,C...与最后的F相比如果重叠度大于IOU的阈值则扔掉原本
设定A和C与F进行重叠比较，如果大于IOU设定阈值则排除AC保留F
顾名思义，保留最大的，抑制较小的
'''

'''
参数设定：
y_pred_decoded（list变量）:就是一批输入框，N个输入框，每个框为一个二维numpy的数组，比如有k个框，那么此参数表示为[k,6]，参数6代表[class_id, score, xmin, xmax, ymin, ymax]向量。
iou_threshold（float变量）：所有的box都会进行相似度度量，如果相似度大于设定的IOU阈值，则从本地列表中删除该框，默认为0.45
 coords（str变量）：设定y_pred_decoded的输入格式
'''
    y_pred_decoded_nms = [] #候选框
    for batch_item in y_pred_decoded:
        boxes_left = np.copy(batch_item) #先将当前的box进行复制
        maxima = []  #存非极大抑制box的临时数组
        while boxes_left.shape[0] > 0:#看box个数是否为0
            maximum_index = np.argmax(boxes_left[:,1]) 
            maximum_box = np.copy(boxes_left[maximum_index]) 
            maxima.append(maximum_box) #存储极大框
            boxes_left = np.delete(boxes_left, maximum_index, axis=0) #删除极大框
            if boxes_left.shape[0] == 0: break