def inference(image, keep_prob): """ Semantic segmentation network definition # 语义分割网络定义 :param image: input image. Should have values in range 0-255 :param keep_prob: :return: """ # 获取预训练网络VGG print("setting up vgg initialized conv layers ...") # model_dir Model_zoo/ # MODEL_URL 下载VGG19网址 model_data = utils.get_model_data(FLAGS.model_dir, MODEL_URL) # 返回VGG19模型中内容 mean = model_data['normalization'][0][0][0] # 获得图像均值 mean_pixel = np.mean(mean, axis=(0, 1)) # RGB weights = np.squeeze(model_data['layers']) # 压缩VGG网络中参数,把维度是1的维度去掉 剩下的就是权重 processed_image = utils.process_image(image, mean_pixel) # 图像减均值 with tf.variable_scope("inference"): # 命名作用域 是inference image_net = vgg_net(weights, processed_image) # 传入权重参数和预测图像,获得所有层输出结果 # conv_final_layer = image_net["conv5_3"] # 获得输出结果 conv_final_layer = image_net["relu4_4"] w5_0 = utils.weight_variable([3, 3, 512, 512], name="W5_0") #取消pool4降采样操作,改成3*3/s1 b5_0 = utils.bias_variable([512], name="b5_0") conv5_0 = utils.conv2d_strided(conv_final_layer, w5_0, b5_0) w5_1 = utils.weight_variable([3, 3, 512, 512], name="W5_1") b5_1 = utils.bias_variable([512], name="b5_1") conv5_1 = utils.conv2d_atrous_2(conv5_0, w5_1, b5_1) w5_2 = utils.weight_variable([3, 3, 512, 512], name="W5_2") #将第五层的conv5_1,2,3改成2-空洞卷积 b5_2 = utils.bias_variable([512], name="b5_2") conv5_2 = utils.conv2d_atrous_2(conv5_1, w5_2, b5_2) w5_3 = utils.weight_variable([3, 3, 512, 512], name="W5_3") b5_3 = utils.bias_variable([512], name="b5_3") conv5_3 = utils.conv2d_atrous_2(conv5_2, w5_3, b5_3) # pool5 = utils.max_pool_2x2(conv_final_layer) # /32 缩小32倍 # W6 = utils.weight_variable([7, 7, 512, 4096], name="W6") # 初始化第6层的w b # b6 = utils.bias_variable([4096], name="b6") # conv6 = utils.conv2d_basic(pool5, W6, b6) # relu6 = tf.nn.relu(conv6, name="relu6") # if FLAGS.debug: # utils.add_activation_summary(relu6) # relu_dropout6 = tf.nn.dropout(relu6, keep_prob=keep_prob) w6_0 = utils.weight_variable([3, 3, 512, 4096], name="W6_0") # 取消pool5降采样操作,改成3*3/s1 b6_0 = utils.bias_variable([4096], name="b6_0") conv6_0 = utils.conv2d_strided(conv5_3, w6_0, b6_0) w6 = utils.weight_variable([3, 3, 4096, 4096], name="W7") b6 = utils.bias_variable([4096], name="b6") #第6层为4-空洞卷积 conv6 = utils.conv2d_atrous_4(conv6_0, w6, b6) relu6 = tf.nn.relu(conv6, name="relu6") if FLAGS.debug: utils.add_activation_summary(relu6) relu_dropout6 = tf.nn.dropout(relu6, keep_prob=keep_prob) # W7 = utils.weight_variable([1, 1, 4096, 4096], name="W7") # 第7层卷积层 # b7 = utils.bias_variable([4096], name="b7") # conv7 = utils.conv2d_basic(relu_dropout6, W7, b7) # relu7 = tf.nn.relu(conv7, name="relu7") # if FLAGS.debug: # utils.add_activation_summary(relu7) # relu_dropout7 = tf.nn.dropout(relu7, keep_prob=keep_prob) w7 = utils.weight_variable([1, 1, 4096, 4096], name="w7") b7 = utils.bias_variable([4096], name="b7") #第7层为4—空洞卷积 conv7 = utils.conv2d_atrous_4(relu_dropout6, w7, b7) relu7 = tf.nn.relu(conv7, name="relu6") if FLAGS.debug: utils.add_activation_summary(relu7) relu_dropout7 = tf.nn.dropout(relu7, keep_prob=keep_prob) # W8 = utils.weight_variable([1, 1, 4096, NUM_OF_CLASSESS], name="W8") # b8 = utils.bias_variable([NUM_OF_CLASSESS], name="b8") # conv8 = utils.conv2d_basic(relu_dropout7, W8, b8) # 第8层卷积层 分类151类 # # annotation_pred1 = tf.argmax(conv8, dimension=3, name="prediction1") w8 = utils.weight_variable([1, 1, 4096, NUM_OF_CLASSESS], name="W8") b8 = utils.bias_variable([NUM_OF_CLASSESS], name="b8") # 第8层为4—空洞卷积 conv8 = utils.conv2d_atrous_4(relu_dropout7, w8, b8) # conv8 = utils.max_pool_2x2(conv8) print(conv8.shape) # now to upscale to actual image size # deconv_shape1 = image_net["pool4"].get_shape() # 将pool4 1/16结果尺寸拿出来 做融合 [b,h,w,c] # # 定义反卷积层的 W,B [H, W, OUTC, INC] 输出个数为pool4层通道个数,输入为conv8通道个数 # # 扩大两倍 所以stride = 2 kernel_size = 4 # W_t1 = utils.weight_variable([4, 4, deconv_shape1[3].value, NUM_OF_CLASSESS], name="W_t1") # b_t1 = utils.bias_variable([deconv_shape1[3].value], name="b_t1") # # 输入为conv8特征图,使得其特征图大小扩大两倍,并且特征图个数变为pool4的通道数 # conv_t1 = utils.conv2d_transpose_strided(conv8, W_t1, b_t1, output_shape=tf.shape(image_net["pool4"])) # fuse_1 = tf.add(conv_t1, image_net["pool4"], name="fuse_1") # 进行融合 逐像素相加 deconv_shape1 = image_net["pool4"].get_shape( ) # 将pool4 1/16结果尺寸拿出来 做融合 [b,h,w,c] # 定义反卷积层的 W,B [H, W, OUTC, INC] 输出个数为pool4层通道个数,输入为conv8通道个数 # 扩大两倍 所以stride = 2 kernel_size = 4 W_t1 = utils.weight_variable( [4, 4, deconv_shape1[3].value, NUM_OF_CLASSESS], name="W_t1") b_t1 = utils.bias_variable([deconv_shape1[3].value], name="b_t1") # 输入为conv8特征图,使得其特征图大小扩大两倍,并且特征图个数变为pool4的通道数 conv_t1 = utils.conv2d_transpose_strided(conv8, W_t1, b_t1, output_shape=tf.shape( image_net["pool4"])) fuse_1 = tf.add(conv_t1, image_net["pool4"], name="fuse_1") # 进行融合 逐像素相加 # 获得pool3尺寸 是原图大小的1/8 deconv_shape2 = image_net["pool3"].get_shape() # 输出通道数为pool3通道数, 输入通道数为pool4通道数 W_t2 = utils.weight_variable( [4, 4, deconv_shape2[3].value, deconv_shape1[3].value], name="W_t2") b_t2 = utils.bias_variable([deconv_shape2[3].value], name="b_t2") # 将上一层融合结果fuse_1在扩大两倍,输出尺寸和pool3相同 conv_t2 = utils.conv2d_transpose_strided(fuse_1, W_t2, b_t2, output_shape=tf.shape( image_net["pool3"])) # 融合操作deconv(fuse_1) + pool3 fuse_2 = tf.add(conv_t2, image_net["pool3"], name="fuse_2") shape = tf.shape(image) # 获得原始图像大小 # 堆叠列表,反卷积输出尺寸,[b,原图H,原图W,类别个数] deconv_shape3 = tf.stack( [shape[0], shape[1], shape[2], NUM_OF_CLASSESS]) # 建立反卷积w[8倍扩大需要ks=16, 输出通道数为类别个数, 输入通道数pool3通道数] W_t3 = utils.weight_variable( [16, 16, deconv_shape2[3].value, NUM_OF_CLASSESS], name="W_t3") b_t3 = utils.bias_variable([NUM_OF_CLASSESS], name="b_t3") # 反卷积,fuse_2反卷积,输出尺寸为 [b,原图H,原图W,类别个数] conv_t3 = utils.conv2d_transpose_strided(fuse_2, W_t3, b_t3, output_shape=deconv_shape3, stride=8) # deconv_shape3 = tf.stack([shape[0], shape[1], shape[2], NUM_OF_CLASSESS]) # # 建立反卷积w[8倍扩大需要ks=16, 输出通道数为类别个数, 输入通道数pool3通道数] # W_t1 = utils.weight_variable([16, 16, NUM_OF_CLASSESS, NUM_OF_CLASSESS], name="W_t1") ##反卷积生成原图大小 # b_t1 = utils.bias_variable([NUM_OF_CLASSESS], name="b_t1") # # 反卷积,fuse_2反卷积,输出尺寸为 [b,原图H,原图W,类别个数] # conv_t1 = utils.conv2d_transpose_strided(conv8, W_t1, b_t1, output_shape=deconv_shape3, stride=8) # 目前conv_t3的形式为size为和原始图像相同的size,通道数与分类数相同 # 这句我的理解是对于每个像素位置,根据第3维度(通道数)通过argmax能计算出这个像素点属于哪个分类 # 也就是对于每个像素而言,NUM_OF_CLASSESS个通道中哪个数值最大,这个像素就属于哪个分类 # 每个像素点有21个值,哪个值最大就属于那一类 # 返回一张图,每一个点对于其来别信息shape=[b,h,w] # annotation_pred = tf.argmax(conv_t1, dimension=3, name="prediction") annotation_pred = tf.argmax(conv_t3, axis=3, name="prediction") # 从第三维度扩展 形成[b,h,w,c] 其中c=1, conv_t3最后具有21深度的特征图 # return tf.expand_dims(annotation_pred, dim=3), conv_t1 return tf.expand_dims(annotation_pred, axis=3), conv_t3