diff --git a/core/general-server/op/general_detection_op.cpp b/core/general-server/op/general_detection_op.cpp
index 46f5ddf1b508681661b69c60a25b6d7d000e6d4e..b62a2d2544e12d493033cf1bb8e6606d72f614d3 100644
--- a/core/general-server/op/general_detection_op.cpp
+++ b/core/general-server/op/general_detection_op.cpp
@@ -191,42 +191,64 @@ int GeneralDetectionOp::inference() {
 
     boxes = post_processor_.FilterTagDetRes(boxes, ratio_h, ratio_w, srcimg);
 
-    for (int i = boxes.size() - 1; i >= 0; i--) {
-      crop_img = GetRotateCropImage(img, boxes[i]);
-
-      float wh_ratio = float(crop_img.cols) / float(crop_img.rows);
+    float max_wh_ratio = 0.0f;
+    std::vector<cv::Mat> crop_imgs;
+    std::vector<cv::Mat> resize_imgs;
+    int max_resize_w = 0;
+    int max_resize_h = 0;
+    int box_num = boxes.size();
+    std::vector<std::vector<float>> output_rec;
+    for (int i = 0; i < box_num; ++i) {
+      cv::Mat line_img = GetRotateCropImage(img, boxes[i]);
+      float wh_ratio = float(line_img.cols) / float(line_img.rows);
+      max_wh_ratio = max_wh_ratio > wh_ratio ? max_wh_ratio : wh_ratio;
+      crop_imgs.push_back(line_img);
+    }
 
+    for (int i = 0; i < box_num; ++i) {
+      cv::Mat resize_img;
+      crop_img = crop_imgs[i];
       this->resize_op_rec.Run(
-          crop_img, resize_img_rec, wh_ratio, this->use_tensorrt_);
+          crop_img, resize_img, max_wh_ratio, this->use_tensorrt_);
 
       this->normalize_op_.Run(
-          &resize_img_rec, this->mean_rec, this->scale_rec, this->is_scale_);
-
-      std::vector<float> output_rec(
-          1 * 3 * resize_img_rec.rows * resize_img_rec.cols, 0.0f);
-
-      this->permute_op_.Run(&resize_img_rec, output_rec.data());
-
-      // Inference.
-      output_shape = {1, 3, resize_img_rec.rows, resize_img_rec.cols};
-      out_num = std::accumulate(
-          output_shape.begin(), output_shape.end(), 1, std::multiplies<int>());
-      databuf_size_out = out_num * sizeof(float);
-      databuf_data_out = MempoolWrapper::instance().malloc(databuf_size_out);
-      if (!databuf_data_out) {
-        LOG(ERROR) << "Malloc failed, size: " << databuf_size_out;
-        return -1;
-      }
-      memcpy(databuf_data_out, output_rec.data(), databuf_size_out);
-      databuf_char_out = reinterpret_cast<char*>(databuf_data_out);
-      paddle::PaddleBuf paddleBuf(databuf_char_out, databuf_size_out);
-      paddle::PaddleTensor tensor_out;
-      tensor_out.name = "image";
-      tensor_out.dtype = paddle::PaddleDType::FLOAT32;
-      tensor_out.shape = {1, 3, resize_img_rec.rows, resize_img_rec.cols};
-      tensor_out.data = paddleBuf;
-      out->push_back(tensor_out);
+          &resize_img, this->mean_rec, this->scale_rec, this->is_scale_);
+
+      max_resize_w = std::max(max_resize_w, resize_img.cols);
+      max_resize_h = std::max(max_resize_h, resize_img.rows);
+      resize_imgs.push_back(resize_img);
+    }
+    int buf_size = 3 * max_resize_h * max_resize_w;
+    output_rec = std::vector<std::vector<float>>(box_num,
+                     std::vector<float>(buf_size, 0.0f));
+    for (int i = 0; i < box_num; ++i) {
+      resize_img_rec = resize_imgs[i];
+
+      this->permute_op_.Run(&resize_img_rec, output_rec[i].data());
+    }
+
+    // Inference.
+    output_shape = {box_num, 3, max_resize_h, max_resize_w};
+    out_num = std::accumulate(
+        output_shape.begin(), output_shape.end(), 1, std::multiplies<int>());
+    databuf_size_out = out_num * sizeof(float);
+    databuf_data_out = MempoolWrapper::instance().malloc(databuf_size_out);
+    if (!databuf_data_out) {
+      LOG(ERROR) << "Malloc failed, size: " << databuf_size_out;
+      return -1;
+    }
+    int offset = buf_size * sizeof(float);
+    for (int i = 0; i < box_num; ++i) {
+      memcpy(databuf_data_out + i * offset, output_rec[i].data(), offset);
     }
+    databuf_char_out = reinterpret_cast<char*>(databuf_data_out);
+    paddle::PaddleBuf paddleBuf(databuf_char_out, databuf_size_out);
+    paddle::PaddleTensor tensor_out;
+    tensor_out.name = "image";
+    tensor_out.dtype = paddle::PaddleDType::FLOAT32;
+    tensor_out.shape = output_shape;
+    tensor_out.data = paddleBuf;
+    out->push_back(tensor_out);
   }
   out->erase(out->begin(), out->begin() + infer_outnum);
 
diff --git a/core/general-server/op/general_detection_op.h b/core/general-server/op/general_detection_op.h
index 272ed5ff40575d42ac3058ad1824285925fc252c..2cc027f5ed761f2d040c0c1858e81cb70a93fcb0 100644
--- a/core/general-server/op/general_detection_op.h
+++ b/core/general-server/op/general_detection_op.h
@@ -63,7 +63,7 @@ class GeneralDetectionOp
 
     double det_db_thresh_ = 0.3;
     double det_db_box_thresh_ = 0.5;
-    double det_db_unclip_ratio_ = 2.0;
+    double det_db_unclip_ratio_ = 1.5;
 
     std::vector<float> mean_det = {0.485f, 0.456f, 0.406f};
     std::vector<float> scale_det = {1 / 0.229f, 1 / 0.224f, 1 / 0.225f};
diff --git a/core/predictor/tools/ocrtools/preprocess_op.cpp b/core/predictor/tools/ocrtools/preprocess_op.cpp
index ab69e4d23abbcbebfbfb5c453fbca46ff5e51967..045984e4c004f965d52badc8b8a0b8996224ab7c 100644
--- a/core/predictor/tools/ocrtools/preprocess_op.cpp
+++ b/core/predictor/tools/ocrtools/preprocess_op.cpp
@@ -82,14 +82,14 @@ void ResizeImgType0::Run(const cv::Mat &img, cv::Mat &resize_img,
   else if (resize_h / 32 < 1 + 1e-5)
     resize_h = 32;
   else
-    resize_h = (resize_h / 32) * 32;
+    resize_h = (resize_h / 32 - 1) * 32;
 
   if (resize_w % 32 == 0)
     resize_w = resize_w;
   else if (resize_w / 32 < 1 + 1e-5)
     resize_w = 32;
   else
-    resize_w = (resize_w / 32) * 32;
+    resize_w = (resize_w / 32 - 1) * 32;
   if (!use_tensorrt) {
     cv::resize(img, resize_img, cv::Size(resize_w, resize_h));
     ratio_h = float(resize_h) / float(h);