detect fields

lib/pdfium.rb

@@ -25,6 +25,8 @@ class Pdfium
   typedef :pointer, :FPDF_BITMAP
   typedef :pointer, :FPDF_FORMHANDLE
   typedef :pointer, :FPDF_TEXTPAGE
+  typedef :pointer, :FPDF_PAGEOBJECT
+  typedef :pointer, :FPDF_PATHSEGMENT
 
   MAX_SIZE = 32_767
 
@@ -37,6 +39,9 @@ class Pdfium
   FPDF_RENDER_FORCEHALFTONE = 0x400
   FPDF_PRINTING = 0x800
 
+  TextNode = Struct.new(:content, :x, :y, :w, :h, keyword_init: true)
+  LineNode = Struct.new(:x, :y, :w, :h, :tilt, keyword_init: true)
+
   # rubocop:disable Naming/ClassAndModuleCamelCase
   class FPDF_LIBRARY_CONFIG < FFI::Struct
     layout :version, :int,
@@ -77,6 +82,37 @@ class Pdfium
   attach_function :FPDFText_ClosePage, [:FPDF_TEXTPAGE], :void
   attach_function :FPDFText_CountChars, [:FPDF_TEXTPAGE], :int
   attach_function :FPDFText_GetText, %i[FPDF_TEXTPAGE int int pointer], :int
+  attach_function :FPDFText_GetUnicode, %i[FPDF_TEXTPAGE int], :uint
+  attach_function :FPDFText_GetCharBox, %i[FPDF_TEXTPAGE int pointer pointer pointer pointer], :int
+  attach_function :FPDFText_GetCharOrigin, %i[FPDF_TEXTPAGE int pointer pointer], :int
+  attach_function :FPDFText_GetCharIndexAtPos, %i[FPDF_TEXTPAGE double double double double], :int
+  attach_function :FPDFText_CountRects, %i[FPDF_TEXTPAGE int int], :int
+  attach_function :FPDFText_GetRect, %i[FPDF_TEXTPAGE int pointer pointer pointer pointer], :int
+  attach_function :FPDFText_GetFontSize, %i[FPDF_TEXTPAGE int], :double
+
+  # Page object functions for extracting paths/lines
+  attach_function :FPDFPage_CountObjects, [:FPDF_PAGE], :int
+  attach_function :FPDFPage_GetObject, %i[FPDF_PAGE int], :FPDF_PAGEOBJECT
+  attach_function :FPDFPageObj_GetType, [:FPDF_PAGEOBJECT], :int
+  attach_function :FPDFPageObj_GetBounds, %i[FPDF_PAGEOBJECT pointer pointer pointer pointer], :int
+  attach_function :FPDFPath_CountSegments, [:FPDF_PAGEOBJECT], :int
+  attach_function :FPDFPath_GetPathSegment, %i[FPDF_PAGEOBJECT int], :FPDF_PATHSEGMENT
+  attach_function :FPDFPathSegment_GetType, [:FPDF_PATHSEGMENT], :int
+  attach_function :FPDFPathSegment_GetPoint, %i[FPDF_PATHSEGMENT pointer pointer], :int
+
+  # Page object types
+  FPDF_PAGEOBJ_UNKNOWN = 0
+  FPDF_PAGEOBJ_TEXT = 1
+  FPDF_PAGEOBJ_PATH = 2
+  FPDF_PAGEOBJ_IMAGE = 3
+  FPDF_PAGEOBJ_SHADING = 4
+  FPDF_PAGEOBJ_FORM = 5
+
+  # Path segment types
+  FPDF_SEGMENT_UNKNOWN = -1
+  FPDF_SEGMENT_LINETO = 0
+  FPDF_SEGMENT_BEZIERTO = 1
+  FPDF_SEGMENT_MOVETO = 2
 
   typedef :int, :FPDF_BOOL
   typedef :pointer, :IPDF_JSPLATFORM
@@ -157,6 +193,7 @@ class Pdfium
     raise PdfiumError, "#{context_message}: #{error_message(error_code)} (Code: #{error_code})"
   end
 
+  # rubocop:disable Metrics
   class Document
     attr_reader :document_ptr, :form_handle
 
@@ -386,6 +423,128 @@ class Pdfium
       Pdfium.FPDFText_ClosePage(text_page) if text_page && !text_page.null?
     end
 
+    def text_nodes
+      return @text_nodes if @text_nodes
+
+      text_page = Pdfium.FPDFText_LoadPage(page_ptr)
+      char_count = Pdfium.FPDFText_CountChars(text_page)
+
+      @text_nodes = []
+
+      return @text_nodes if char_count.zero?
+
+      char_count.times do |i|
+        unicode = Pdfium.FPDFText_GetUnicode(text_page, i)
+
+        char = [unicode].pack('U*')
+
+        left_ptr = FFI::MemoryPointer.new(:double)
+        right_ptr = FFI::MemoryPointer.new(:double)
+        bottom_ptr = FFI::MemoryPointer.new(:double)
+        top_ptr = FFI::MemoryPointer.new(:double)
+
+        result = Pdfium.FPDFText_GetCharBox(text_page, i, left_ptr, right_ptr, bottom_ptr, top_ptr)
+
+        next if result.zero?
+
+        left = left_ptr.read_double
+        right = right_ptr.read_double
+
+        origin_x_ptr = FFI::MemoryPointer.new(:double)
+        origin_y_ptr = FFI::MemoryPointer.new(:double)
+
+        Pdfium.FPDFText_GetCharOrigin(text_page, i, origin_x_ptr, origin_y_ptr)
+
+        origin_y = origin_y_ptr.read_double
+
+        font_size = Pdfium.FPDFText_GetFontSize(text_page, i)
+        font_size = 8 if font_size == 1
+
+        abs_x = left
+        abs_y = height - origin_y - (font_size * 0.8)
+        abs_width = right - left
+        abs_height = font_size
+
+        x = abs_x / width
+        y = abs_y / height
+        node_width = abs_width / width
+        node_height = abs_height / height
+
+        @text_nodes << TextNode.new(content: char, x: x, y: y, w: node_width, h: node_height)
+      end
+
+      @text_nodes = @text_nodes.sort { |a, b| a.y == b.y ? a.x <=> b.x : a.y <=> b.y }
+    ensure
+      Pdfium.FPDFText_ClosePage(text_page) if text_page && !text_page.null?
+    end
+
+    def line_nodes
+      return @line_nodes if @line_nodes
+
+      ensure_not_closed!
+
+      @line_nodes = []
+
+      object_count = Pdfium.FPDFPage_CountObjects(page_ptr)
+
+      return @line_nodes if object_count.zero?
+
+      object_count.times do |i|
+        page_object = Pdfium.FPDFPage_GetObject(page_ptr, i)
+
+        next if page_object.null?
+
+        obj_type = Pdfium.FPDFPageObj_GetType(page_object)
+
+        next unless obj_type == Pdfium::FPDF_PAGEOBJ_PATH
+
+        left_ptr = FFI::MemoryPointer.new(:float)
+        bottom_ptr = FFI::MemoryPointer.new(:float)
+        right_ptr = FFI::MemoryPointer.new(:float)
+        top_ptr = FFI::MemoryPointer.new(:float)
+
+        Pdfium.FPDFPageObj_GetBounds(page_object, left_ptr, bottom_ptr, right_ptr, top_ptr)
+
+        obj_left = left_ptr.read_float
+        obj_bottom = bottom_ptr.read_float
+        obj_right = right_ptr.read_float
+        obj_top = top_ptr.read_float
+
+        obj_width = obj_right - obj_left
+        obj_height = obj_top - obj_bottom
+
+        next if obj_width < 1 && obj_height < 1
+
+        segment_count = Pdfium.FPDFPath_CountSegments(page_object)
+
+        next if segment_count < 2
+
+        next unless segment_count <= 10 && (obj_height < 10 || obj_width < 10)
+
+        if obj_width > obj_height && obj_height < 10
+          tilt = 0
+        elsif obj_height > obj_width && obj_width < 10
+          tilt = 90
+        else
+          next
+        end
+
+        x = obj_left
+        y = obj_bottom
+        w = obj_width
+        h = obj_height
+
+        norm_x = x / width
+        norm_y = (height - y - h) / height
+        norm_w = w / width
+        norm_h = h / height
+
+        @line_nodes << LineNode.new(x: norm_x, y: norm_y, w: norm_w, h: norm_h, tilt: tilt)
+      end
+
+      @line_nodes = @line_nodes.sort { |a, b| a.y == b.y ? a.x <=> b.x : a.y <=> b.y }
+    end
+
     def close
       return if closed?
 
@@ -445,4 +604,5 @@ class Pdfium
   at_exit do
     cleanup_library
   end
+  # rubocop:enable Metrics
 end

lib/templates/detect_fields.rb

@@ -0,0 +1,264 @@
+# frozen_string_literal: true
+
+module Templates
+  module DetectFields
+    module_function
+
+    TextFieldBox = Struct.new(:x, :y, :w, :h, keyword_init: true)
+
+    # rubocop:disable Metrics
+    def call(io, attachment: nil, confidence: 0.3, temperature: 1,
+             nms: 0.1, split_page: false, aspect_ratio: true, padding: 20, &)
+      if attachment&.image?
+        process_image_attachment(io, attachment:, confidence:, nms:, split_page:,
+                                     temperature:, aspect_ratio:, padding:, &)
+      else
+        process_pdf_attachment(io, attachment:, confidence:, nms:, split_page:,
+                                   temperature:, aspect_ratio:, padding:, &)
+      end
+    end
+
+    def process_image_attachment(io, attachment:, confidence:, nms:, temperature: 1,
+                                 split_page: false, aspect_ratio: false, padding: nil)
+      image = Vips::Image.new_from_buffer(io.read, '')
+
+      fields = Templates::ImageToFields.call(image, confidence:, nms:, split_page:,
+                                                    temperature:, aspect_ratio:, padding:)
+
+      fields = fields.map do |f|
+        {
+          uuid: SecureRandom.uuid,
+          type: f.type,
+          required: true,
+          preferences: {},
+          areas: [{
+            x: f.x,
+            y: f.y,
+            w: f.w,
+            h: f.h,
+            page: 0,
+            attachment_uuid: attachment&.uuid
+          }]
+        }
+      end
+
+      yield [attachment&.uuid, 0, fields] if block_given?
+
+      fields
+    end
+
+    def process_pdf_attachment(io, attachment:, confidence:, nms:, temperature: 1,
+                               split_page: false, aspect_ratio: false, padding: nil)
+      doc = Pdfium::Document.open_bytes(io.read)
+
+      doc.page_count.times.flat_map do |page_number|
+        page = doc.get_page(page_number)
+
+        data, width, height = page.render_to_bitmap(width: ImageToFields::RESOLUTION * 1.5)
+
+        image = Vips::Image.new_from_memory(data, width, height, 4, :uchar)
+
+        fields = Templates::ImageToFields.call(image, confidence: 0.05, nms:, split_page:,
+                                                      temperature:, aspect_ratio:, padding:)
+
+        text_fields = extract_text_fields_from_page(page)
+        line_fields = extract_line_fields_from_page(page)
+
+        fields = increase_confidence_for_overlapping_fields(fields, text_fields)
+        fields = increase_confidence_for_overlapping_fields(fields, line_fields)
+
+        fields = fields.filter_map do |f|
+          next if f.confidence < confidence
+
+          {
+            uuid: SecureRandom.uuid,
+            type: f.type,
+            required: true,
+            preferences: {},
+            areas: [{
+              x: f.x, y: f.y,
+              w: f.w, h: f.h,
+              page: page_number,
+              attachment_uuid: attachment&.uuid
+            }]
+          }
+        end
+
+        yield [attachment&.uuid, page_number, fields] if block_given?
+
+        fields
+      end
+    end
+
+    def extract_line_fields_from_page(page)
+      line_thickness = 5.0 / page.height
+
+      vertical_lines, all_horizontal_lines = page.line_nodes.partition { |line| line.tilt == 90 }
+
+      horizontal_lines = all_horizontal_lines.reject do |h_line|
+        next true if h_line.w > 0.7 && (h_line.h < 0.1 || h_line.h < 0.9)
+
+        next false if vertical_lines.blank?
+
+        h_x_min = h_line.x
+        h_x_max = h_line.x + h_line.w
+        h_y_avg = h_line.y + (h_line.h / 2)
+
+        vertical_lines.any? do |v_line|
+          v_x_avg = v_line.x + (v_line.w / 2)
+          v_y_min = v_line.y
+          v_y_max = v_line.y + v_line.h
+
+          h_x_min_expanded = h_x_min - line_thickness
+          h_x_max_expanded = h_x_max + line_thickness
+          h_y_min_expanded = h_y_avg - line_thickness
+          h_y_max_expanded = h_y_avg + line_thickness
+
+          v_x_min_expanded = v_x_avg - line_thickness
+          v_x_max_expanded = v_x_avg + line_thickness
+          v_y_min_expanded = v_y_min - line_thickness
+          v_y_max_expanded = v_y_max + line_thickness
+
+          x_overlap = v_x_min_expanded <= h_x_max_expanded && v_x_max_expanded >= h_x_min_expanded
+          y_overlap = h_y_min_expanded <= v_y_max_expanded && h_y_max_expanded >= v_y_min_expanded
+
+          x_overlap && y_overlap
+        end
+      end
+
+      node_index = 0
+
+      horizontal_lines = horizontal_lines.reject do |line|
+        nodes = []
+
+        loop do
+          node = page.text_nodes[node_index += 1]
+
+          break unless node
+
+          break if node.y > line.y
+
+          next if node.x + node.w < line.x || line.x + line.w < node.x ||
+                  node.y + node.h < line.y - node.h || line.y < node.y
+
+          nodes << node
+
+          next if nodes.blank?
+
+          next_node = page.text_nodes[node_index + 1]
+
+          break if next_node.x + next_node.w < line.x || line.x + line.w < next_node.x ||
+                   next_node.y + next_node.h < line.y - next_node.h || line.y < next_node.y
+        end
+
+        next if nodes.blank?
+
+        width = nodes.last.x + nodes.last.w - nodes.first.x
+
+        next true if width > line.w / 2.0
+      end
+
+      horizontal_lines.each do |line|
+        line.h += 4 * line_thickness
+        line.y -= 4 * line_thickness
+      end
+    end
+
+    def extract_text_fields_from_page(page)
+      text_nodes = page.text_nodes
+
+      field_boxes = []
+
+      i = 0
+
+      while i < text_nodes.length
+        node = text_nodes[i]
+
+        next i += 1 if node.content != '_'
+
+        x1 = node.x
+        y1 = node.y
+        x2 = node.x + node.w
+        y2 = node.y + node.h
+
+        underscore_count = 1
+
+        j = i + 1
+
+        while j < text_nodes.length
+          next_node = text_nodes[j]
+
+          break unless next_node.content == '_'
+
+          distance = next_node.x - x2
+          height_diff = (next_node.y - y1).abs
+
+          break if distance > 0.02 || height_diff > node.h * 0.5
+
+          underscore_count += 1
+          next_x2 = next_node.x + next_node.w
+          next_y2 = next_node.y + next_node.h
+
+          x2 = next_x2
+          y2 = [y2, next_y2].max
+          y1 = [y1, next_node.y].min
+
+          j += 1
+        end
+
+        field_boxes << TextFieldBox.new(x: x1, y: y1, w: x2 - x1, h: y2 - y1) if underscore_count >= 2
+
+        i = j
+      end
+
+      field_boxes
+    end
+
+    def calculate_iou(box1, box2)
+      x1 = [box1.x, box2.x].max
+      y1 = [box1.y, box2.y].max
+      x2 = [box1.x + box1.w, box2.x + box2.w].min
+      y2 = [box1.y + box1.h, box2.y + box2.h].min
+
+      intersection_width = [0, x2 - x1].max
+      intersection_height = [0, y2 - y1].max
+      intersection_area = intersection_width * intersection_height
+
+      return 0.0 if intersection_area.zero?
+
+      box1_area = box1.w * box1.h
+      box2_area = box2.w * box2.h
+      union_area = box1_area + box2_area - intersection_area
+
+      intersection_area / union_area
+    end
+
+    def boxes_overlap?(box1, box2)
+      !(box1.x + box1.w < box2.x || box2.x + box2.w < box1.x ||
+        box1.y + box1.h < box2.y || box2.y + box2.h < box1.y)
+    end
+
+    def increase_confidence_for_overlapping_fields(image_fields, text_fields, by: 1.0)
+      return image_fields if text_fields.blank?
+
+      image_fields.map do |image_field|
+        next if image_field.type != 'text'
+
+        field_bottom = image_field.y + image_field.h
+
+        text_fields.each do |text_field|
+          break if text_field.y > field_bottom
+
+          next if text_field.y + text_field.h < image_field.y
+
+          next unless boxes_overlap?(image_field, text_field) && calculate_iou(image_field, text_field) > 0.5
+
+          break image_field.confidence += by
+        end
+      end
+
+      image_fields
+    end
+    # rubocop:enable Metrics
+  end
+end

lib/templates/image_to_fields.rb

@@ -0,0 +1,331 @@
+# frozen_string_literal: true
+
+module Templates
+  module ImageToFields
+    module_function
+
+    Field = Struct.new(:type, :x, :y, :w, :h, :confidence, keyword_init: true)
+
+    MODEL_PATH = Rails.root.join('tmp/model_704_int8.onnx')
+
+    RESOLUTION = 704
+
+    ID_TO_CLASS = %w[text checkbox].freeze
+
+    MEAN = [0.485, 0.456, 0.406].freeze
+    STD = [0.229, 0.224, 0.225].freeze
+
+    CPU_THREADS = Etc.nprocessors
+
+    # rubocop:disable Metrics
+    def call(image, confidence: 0.3, nms: 0.1, temperature: 1,
+             split_page: false, aspect_ratio: true, padding: nil)
+      base_image = image.extract_band(0, n: 3)
+
+      trimmed_base, base_offset_x, base_offset_y = trim_image_with_padding(base_image, padding)
+
+      if split_page && image.height > image.width
+        half_h = trimmed_base.height / 2
+        top_h = half_h
+        bottom_h = trimmed_base.height - half_h
+
+        regions = [
+          { img: trimmed_base.crop(0, 0, trimmed_base.width, top_h), offset_y: 0 },
+          { img: trimmed_base.crop(0, top_h, trimmed_base.width, bottom_h), offset_y: top_h }
+        ]
+
+        detections = { xyxy: Numo::SFloat[], confidence: Numo::SFloat[], class_id: Numo::Int32[] }
+
+        detections = regions.reduce(detections) do |acc, r|
+          next detections if r[:img].height <= 0 || r[:img].width <= 0
+
+          input_tensor, transform_info = preprocess_image(r[:img], RESOLUTION, aspect_ratio:)
+
+          transform_info[:trim_offset_x] = base_offset_x
+          transform_info[:trim_offset_y] = base_offset_y + r[:offset_y]
+
+          outputs = model.predict({ 'input' => input_tensor })
+
+          postprocess_outputs(outputs, transform_info, acc, confidence:, temperature:)
+        end
+      else
+        input_tensor, transform_info = preprocess_image(trimmed_base, RESOLUTION, aspect_ratio:)
+
+        transform_info[:trim_offset_x] = base_offset_x
+        transform_info[:trim_offset_y] = base_offset_y
+
+        outputs = model.predict({ 'input' => input_tensor })
+
+        detections = postprocess_outputs(outputs, transform_info, confidence:, temperature:)
+      end
+
+      detections = apply_nms(detections, nms)
+
+      fields = Array.new(detections[:xyxy].shape[0]) do |i|
+        x1 = detections[:xyxy][i, 0]
+        y1 = detections[:xyxy][i, 1]
+        x2 = detections[:xyxy][i, 2]
+        y2 = detections[:xyxy][i, 3]
+
+        class_id = detections[:class_id][i].to_i
+
+        confidence = detections[:confidence][i]
+
+        x0_norm = x1 / image.width.to_f
+        y0_norm = y1 / image.height.to_f
+        x1_norm = x2 / image.width.to_f
+        y1_norm = y2 / image.height.to_f
+
+        type_name = ID_TO_CLASS[class_id]
+
+        Field.new(
+          type: type_name,
+          x: x0_norm,
+          y: y0_norm,
+          w: (x1_norm - x0_norm),
+          h: (y1_norm - y0_norm),
+          confidence:
+        )
+      end
+
+      sort_fields(fields, y_threshold: 10.0 / image.height)
+    end
+
+    def trim_image_with_padding(image, padding = 0)
+      return [image, 0, 0] if padding.nil?
+
+      left, top, trim_width, trim_height = image.find_trim(threshold: 10, background: [255, 255, 255])
+
+      padded_left = [left - padding, 0].max
+      padded_top = [top - padding, 0].max
+      padded_right = [left + trim_width + padding, image.width].min
+      padded_bottom = [top + trim_height + padding, image.height].min
+
+      width = padded_right - padded_left
+      height = padded_bottom - padded_top
+
+      trimmed_image = image.crop(padded_left, padded_top, width, height)
+
+      [trimmed_image, padded_left, padded_top]
+    end
+
+    def preprocess_image(image, resolution, aspect_ratio: false)
+      scale_x = resolution.to_f / image.width
+      scale_y = resolution.to_f / image.height
+
+      if aspect_ratio
+        scale = [scale_x, scale_y].min
+
+        new_width = (image.width * scale).round
+        new_height = (image.height * scale).round
+
+        resized = image.resize(scale, vscale: scale, kernel: :lanczos3)
+
+        pad_x = ((resolution - new_width) / 2.0).round
+        pad_y = ((resolution - new_height) / 2.0).round
+
+        image = resized.embed(pad_x, pad_y, resolution, resolution, background: [255, 255, 255])
+
+        transform_info = { scale_x: scale, scale_y: scale, pad_x: pad_x, pad_y: pad_y }
+      else
+        image = image.resize(scale_x, vscale: scale_y, kernel: :lanczos3)
+
+        transform_info = { scale_x: scale_x, scale_y: scale_y, pad_x: 0, pad_y: 0 }
+      end
+
+      image /= 255.0
+
+      image = (image - MEAN) / STD
+
+      pixel_data = image.write_to_memory
+
+      img_array = Numo::SFloat.from_binary(pixel_data, [resolution, resolution, 3])
+
+      img_array = img_array.transpose(2, 0, 1)
+
+      [img_array.reshape(1, 3, resolution, resolution), transform_info]
+    end
+
+    def nms(boxes, scores, iou_threshold = 0.5)
+      return Numo::Int32[] if boxes.shape[0].zero?
+
+      x1 = boxes[true, 0]
+      y1 = boxes[true, 1]
+      x2 = boxes[true, 2]
+      y2 = boxes[true, 3]
+
+      areas = (x2 - x1) * (y2 - y1)
+      order = scores.sort_index.reverse
+
+      keep = []
+
+      while order.size.positive?
+        i = order[0]
+        keep << i
+
+        break if order.size == 1
+
+        xx1 = Numo::SFloat.maximum(x1[i], x1[order[1..]])
+        yy1 = Numo::SFloat.maximum(y1[i], y1[order[1..]])
+        xx2 = Numo::SFloat.minimum(x2[i], x2[order[1..]])
+        yy2 = Numo::SFloat.minimum(y2[i], y2[order[1..]])
+
+        w = Numo::SFloat.maximum(0.0, xx2 - xx1)
+        h = Numo::SFloat.maximum(0.0, yy2 - yy1)
+
+        intersection = w * h
+
+        iou = intersection / (areas[i] + areas[order[1..]] - intersection)
+
+        inds = iou.le(iou_threshold).where
+
+        order = order[inds + 1]
+      end
+
+      Numo::Int32.cast(keep)
+    end
+
+    def postprocess_outputs(outputs, transform_info, detections = nil, confidence: 0.3, temperature: 1)
+      boxes = Numo::SFloat.cast(outputs['dets'])
+      logits = Numo::SFloat.cast(outputs['labels'])
+
+      boxes = boxes[0, true, true] # [300, 4]
+      logits = logits[0, true, true] # [300, num_classes]
+
+      scaled_logits = logits / temperature
+
+      probs = 1.0 / (1.0 + Numo::NMath.exp(-scaled_logits))
+
+      scores = probs.max(axis: 1)
+      labels = probs.argmax(axis: 1)
+
+      cx = boxes[true, 0]
+      cy = boxes[true, 1]
+      w = boxes[true, 2]
+      h = boxes[true, 3]
+
+      x1 = cx - (w / 2.0)
+      y1 = cy - (h / 2.0)
+      x2 = cx + (w / 2.0)
+      y2 = cy + (h / 2.0)
+
+      boxes_xyxy = Numo::SFloat.zeros(boxes.shape[0], 4)
+      boxes_xyxy[true, 0] = x1
+      boxes_xyxy[true, 1] = y1
+      boxes_xyxy[true, 2] = x2
+      boxes_xyxy[true, 3] = y2
+
+      boxes_xyxy *= RESOLUTION
+
+      pad_x = transform_info[:pad_x]
+      pad_y = transform_info[:pad_y]
+      boxes_xyxy[true, 0] -= pad_x
+      boxes_xyxy[true, 1] -= pad_y
+      boxes_xyxy[true, 2] -= pad_x
+      boxes_xyxy[true, 3] -= pad_y
+
+      scale_x = transform_info[:scale_x]
+      scale_y = transform_info[:scale_y]
+      boxes_xyxy[true, 0] /= scale_x
+      boxes_xyxy[true, 1] /= scale_y
+      boxes_xyxy[true, 2] /= scale_x
+      boxes_xyxy[true, 3] /= scale_y
+
+      trim_offset_x = transform_info[:trim_offset_x]
+      trim_offset_y = transform_info[:trim_offset_y]
+      boxes_xyxy[true, 0] += trim_offset_x
+      boxes_xyxy[true, 1] += trim_offset_y
+      boxes_xyxy[true, 2] += trim_offset_x
+      boxes_xyxy[true, 3] += trim_offset_y
+
+      keep_mask = scores.gt(confidence)
+
+      keep_indices = keep_mask.where
+
+      if keep_indices.empty?
+        detections || {
+          xyxy: Numo::SFloat[],
+          confidence: Numo::SFloat[],
+          class_id: Numo::Int32[]
+        }
+      else
+        scores = scores[keep_indices]
+        labels = labels[keep_indices]
+        boxes_xyxy = boxes_xyxy[keep_indices, true]
+
+        if detections
+          existing_n = detections[:xyxy].shape[0]
+          new_n = boxes_xyxy.shape[0]
+          total = existing_n + new_n
+
+          xyxy = Numo::SFloat.zeros(total, 4)
+          conf = Numo::SFloat.zeros(total)
+          cls = Numo::Int32.zeros(total)
+
+          if existing_n.positive?
+            xyxy[0...existing_n, true] = detections[:xyxy]
+            conf[0...existing_n] = detections[:confidence]
+            cls[0...existing_n] = detections[:class_id]
+          end
+
+          xyxy[existing_n...(existing_n + new_n), true] = boxes_xyxy
+          conf[existing_n...(existing_n + new_n)] = scores
+          cls[existing_n...(existing_n + new_n)] = Numo::Int32.cast(labels)
+
+          { xyxy: xyxy, confidence: conf, class_id: cls }
+        else
+          {
+            xyxy: boxes_xyxy,
+            confidence: scores,
+            class_id: Numo::Int32.cast(labels)
+          }
+        end
+      end
+    end
+
+    def sort_fields(fields, y_threshold: 0.01)
+      sorted_fields = fields.sort { |a, b| a.y == b.y ? a.x <=> b.x : a.y <=> b.y }
+
+      lines = []
+      current_line = []
+
+      sorted_fields.each do |field|
+        if current_line.blank? || (field.y - current_line.first.y).abs < y_threshold
+          current_line << field
+        else
+          lines << current_line.sort_by(&:x)
+
+          current_line = [field]
+        end
+      end
+
+      lines << current_line.sort_by(&:x) if current_line.present?
+
+      lines.flatten
+    end
+
+    def apply_nms(detections, threshold = 0.5)
+      return detections if detections[:xyxy].shape[0].zero?
+
+      keep_indices = nms(detections[:xyxy], detections[:confidence], threshold)
+
+      {
+        xyxy: detections[:xyxy][keep_indices, true],
+        confidence: detections[:confidence][keep_indices],
+        class_id: detections[:class_id][keep_indices]
+      }
+    end
+
+    def model
+      @model ||= OnnxRuntime::Model.new(
+        MODEL_PATH.to_s,
+        inter_op_num_threads: CPU_THREADS,
+        intra_op_num_threads: CPU_THREADS,
+        enable_mem_pattern: false,
+        enable_cpu_mem_arena: false,
+        providers: ['CPUExecutionProvider']
+      )
+    end
+    # rubocop:enable Metrics
+  end
+end