长春建设集团股份有限公司网站自己如何建设网站聊天室

当前位置： 首页 > news >正文

长春建设集团股份有限公司网站,自己如何建设网站聊天室,北京网站的建立,深圳公司注册下来有哪些资料目录 一、实验目的 二、实验平台 三、实验内容 3.1 绘制锚框 3.2 计算IoU交并比 3.3 绘制所有预测框 3.4 非极大值抑制 3.5 产生候选区域、输出特征图形状 3.6 检测头设计(计算预测框位置和类别) 3.7 目标检测实战#xff1a;基于YOLOv3完成林业病虫害…目录 一、实验目的 二、实验平台 三、实验内容 3.1   绘制锚框 3.2   计算IoU交并比 3.3   绘制所有预测框 3.4   非极大值抑制 3.5   产生候选区域、输出特征图形状 3.6   检测头设计(计算预测框位置和类别) 3.7   目标检测实战基于YOLOv3完成林业病虫害检测 四、总结 一、实验目的 了解python语法了解目标检测的原理实现目标检测YOLOv3并对比不同的结果 二、实验平台 Baidu 飞桨AI Studio 三、实验内容 3.1   绘制锚框 3.1.1 实验代码 import numpy as np import matplotlib.pyplot as plt %matplotlib inline import matplotlib.patches as patches from matplotlib.image import imread import mathdef draw_rectangle(currentAxis, bbox, edgecolor k, facecolor y, fillFalse, linestyle-):rectpatches.Rectangle((bbox[0], bbox[1]), bbox[2]-bbox[0]1, bbox[3]-bbox[1]1, linewidth1,edgecoloredgecolor,facecolorfacecolor,fillfill, linestylelinestyle)currentAxis.add_patch(rect)plt.figure(figsize(10, 10))filename /home/aistudio/work/images/section3/000000086956.jpg im imread(filename) plt.imshow(im)bbox1 [214.29, 325.03, 399.82, 631.37] bbox2 [40.93, 141.1, 226.99, 515.73] bbox3 [247.2, 131.62, 480.0, 639.32]currentAxisplt.gca()draw_rectangle(currentAxis, bbox1, edgecolorr) draw_rectangle(currentAxis, bbox2, edgecolorr) draw_rectangle(currentAxis, bbox3,edgecolorr)def draw_anchor_box(center, length, scales, ratios, img_height, img_width):bboxes []for scale in scales:for ratio in ratios:h length*scale*math.sqrt(ratio)w length*scale/math.sqrt(ratio) x1 max(center[0] - w/2., 0.)y1 max(center[1] - h/2., 0.)x2 min(center[0] w/2. - 1.0, img_width - 1.0)y2 min(center[1] h/2. - 1.0, img_height - 1.0)print(center[0], center[1], w, h)bboxes.append([x1, y1, x2, y2])for bbox in bboxes:draw_rectangle(currentAxis, bbox, edgecolor b)img_height im.shape[0] img_width im.shape[1]         draw_anchor_box([300., 500.], 100., [2.0], [0.5, 1.0, 2.0], img_height, img_width)plt.text(285, 285, G1, colorred, fontsize20) plt.arrow(300, 288, 30, 40, colorred, width0.001, length_includes_headTrue, \head_width5, head_length10, shapefull)plt.text(190, 320, A1, colorblue, fontsize20) plt.arrow(200, 320, 30, 40, colorblue, width0.001, length_includes_headTrue, \head_width5, head_length10, shapefull)plt.text(160, 370, A2, colorblue, fontsize20) plt.arrow(170, 370, 30, 40, colorblue, width0.001, length_includes_headTrue, \head_width5, head_length10, shapefull)plt.text(115, 420, A3, colorblue, fontsize20) plt.arrow(127, 420, 30, 40, colorblue, width0.001, length_includes_headTrue, \head_width5, head_length10, shapefull)plt.show() 3.1.2 绘制锚框实验结果 3.2   计算IoU交并比 3.2.1 实验代码 def box_iou_xyxy(box1, box2):x1min, y1min, x1max, y1max box1[0], box1[1], box1[2], box1[3]s1 (y1max - y1min 1.) * (x1max - x1min 1.)x2min, y2min, x2max, y2max box2[0], box2[1], box2[2], box2[3]s2 (y2max - y2min 1.) * (x2max - x2min 1.)xmin np.maximum(x1min, x2min)ymin np.maximum(y1min, y2min)xmax np.minimum(x1max, x2max)ymax np.minimum(y1max, y2max)inter_h np.maximum(ymax - ymin 1., 0.)inter_w np.maximum(xmax - xmin 1., 0.)intersection inter_h * inter_wunion s1 s2 - intersectioniou intersection / unionreturn ioubbox1 [100., 100., 200., 200.] bbox2 [120., 120., 220., 220.] iou box_iou_xyxy(bbox1, bbox2) print(IoU is {}.format(iou))   3.2.2 实验结果 3.3   绘制所有预测框 3.3.1 实验代码 import numpy as np import matplotlib.pyplot as plt import matplotlib.patches as patches from matplotlib.image import imread import mathdef draw_rectangle(currentAxis, bbox, edgecolor k, facecolor y, fillFalse, linestyle-):rectpatches.Rectangle((bbox[0], bbox[1]), bbox[2]-bbox[0]1, bbox[3]-bbox[1]1, linewidth1,edgecoloredgecolor,facecolorfacecolor,fillfill, linestylelinestyle)currentAxis.add_patch(rect)plt.figure(figsize(10, 10))filename /home/aistudio/000000086956.jpg im imread(filename) plt.imshow(im)currentAxisplt.gca()boxes np.array([[4.21716537e01, 1.28230896e02, 2.26547668e02, 6.00434631e02],[3.18562988e02, 1.23168472e02, 4.79000000e02, 6.05688416e02],[2.62704697e01, 1.39430557e02, 2.20587097e02, 6.38959656e02],[4.24965363e01, 1.42706665e02, 2.25955185e02, 6.35671204e02],[2.37462646e02, 1.35731537e02, 4.79000000e02, 6.31451294e02],[3.19390472e02, 1.29295090e02, 4.79000000e02, 6.33003845e02],[3.28933838e02, 1.22736115e02, 4.79000000e02, 6.39000000e02],[4.44292603e01, 1.70438187e02, 2.26841858e02, 6.39000000e02],[2.17988785e02, 3.02472412e02, 4.06062927e02, 6.29106628e02],[2.00241089e02, 3.23755096e02, 3.96929321e02, 6.36386108e02],[2.14310303e02, 3.23443665e02, 4.06732849e02, 6.35775269e02]])scores np.array([0.5247661 , 0.51759845, 0.86075854, 0.9910175 , 0.39170712,0.9297706 , 0.5115228 , 0.270992  , 0.19087596, 0.64201415, 0.879036])for box in boxes:draw_rectangle(currentAxis, box) 3.3.2 实验结果 3.4   非极大值抑制 3.4.1 实验代码 def nms(bboxes, scores, score_thresh, nms_thresh):inds np.argsort(scores)inds inds[::-1]keep_inds []while(len(inds) 0):cur_ind inds[0]cur_score scores[cur_ind]if cur_score score_thresh:breakkeep Truefor ind in keep_inds:current_box bboxes[cur_ind]remain_box bboxes[ind]iou box_iou_xyxy(current_box, remain_box)if iou nms_thresh:keep Falsebreakif keep:keep_inds.append(cur_ind)inds inds[1:]return np.array(keep_inds)plt.figure(figsize(10, 10)) plt.imshow(im) currentAxisplt.gca() colors [r, g, b, k]inds nms(boxes, scores, score_thresh0.01, nms_thresh0.5) print(inds) for i in range(len(inds)):box boxes[inds[i]]draw_rectangle(currentAxis, box, edgecolorcolors[i]) 3.4.2 实验结果如图5所示 3.5   产生候选区域、输出特征图形状 3.5.1 实验代码 import paddle import paddle.nn.functional as F import numpy as npclass ConvBNLayer(paddle.nn.Layer):def init(self, ch_in, ch_out, kernel_size3, stride1, groups1,padding0, actleaky):super(ConvBNLayer, self).init()self.conv paddle.nn.Conv2D(in_channelsch_in,out_channelsch_out,kernel_sizekernel_size,stridestride,paddingpadding,groupsgroups,weight_attrpaddle.ParamAttr(initializerpaddle.nn.initializer.Normal(0., 0.02)),bias_attrFalse)self.batch_norm paddle.nn.BatchNorm2D(num_featuresch_out,weight_attrpaddle.ParamAttr(initializerpaddle.nn.initializer.Normal(0., 0.02),regularizerpaddle.regularizer.L2Decay(0.)),bias_attrpaddle.ParamAttr(initializerpaddle.nn.initializer.Constant(0.0),regularizerpaddle.regularizer.L2Decay(0.)))self.act actdef forward(self, inputs):out self.conv(inputs)out self.batch_norm(out)if self.act leaky:out F.leaky_relu(xout, negative_slope0.1)return outclass DownSample(paddle.nn.Layer):def init(self,ch_in,ch_out,kernel_size3,stride2,padding1):super(DownSample, self).init()self.conv_bn_layer ConvBNLayer(ch_inch_in,ch_outch_out,kernel_sizekernel_size,stridestride,paddingpadding)self.ch_out ch_outdef forward(self, inputs):out self.conv_bn_layer(inputs)return outclass BasicBlock(paddle.nn.Layer):def init(self, ch_in, ch_out):super(BasicBlock, self).init()self.conv1 ConvBNLayer(ch_inch_in,ch_outch_out,kernel_size1,stride1,padding0)self.conv2 ConvBNLayer(ch_inch_out,ch_outch_out*2,kernel_size3,stride1,padding1)def forward(self, inputs):conv1 self.conv1(inputs)conv2 self.conv2(conv1)out paddle.add(xinputs, yconv2)return outclass LayerWarp(paddle.nn.Layer):def init(self, ch_in, ch_out, count, is_testTrue):super(LayerWarp,self).init()self.basicblock0 BasicBlock(ch_in,ch_out)self.res_out_list []for i in range(1, count):res_out self.add_sublayer(basicblock%d % (i), BasicBlock(ch_out*2,ch_out))self.res_out_list.append(res_out)def forward(self,inputs):y self.basicblock0(inputs)for basic_block_i in self.res_out_list:y basic_block_i(y)return yDarkNet_cfg {53: ([1, 2, 8, 8, 4])}class DarkNet53_conv_body(paddle.nn.Layer):def init(self):super(DarkNet53_conv_body, self).init()self.stages DarkNet_cfg[53]self.stages self.stages[0:5]self.conv0 ConvBNLayer(ch_in3,ch_out32,kernel_size3,stride1,padding1)self.downsample0 DownSample(ch_in32,ch_out32 * 2)self.darknet53_conv_block_list []self.downsample_list []for i, stage in enumerate(self.stages):conv_block self.addsublayer(stage%d % (i),LayerWarp(32(2**(i1)),32(2i),stage))self.darknet53_conv_block_list.append(conv_block)for i in range(len(self.stages) - 1):downsample self.addsublayer(stage%d_downsample % i,DownSample(ch_in32*(2(i1)),ch_out32*(2**(i2))))self.downsample_list.append(downsample)def forward(self,inputs):out self.conv0(inputs)out self.downsample0(out)blocks []for i, conv_block_i in enumerate(self.darknet53_conv_block_list): out conv_block_i(out)blocks.append(out)if i len(self.stages) - 1:out self.downsample_listireturn blocks[-1:-4:-1] import numpy as np backbone DarkNet53_conv_body() x np.random.randn(1, 3, 640, 640).astype(float32) x paddle.to_tensor(x) C0, C1, C2 backbone(x) print(C0.shape, C1.shape, C2.shape) 3.6   检测头设计(计算预测框位置和类别) 3.6.1 对C0进行多次卷积以得到跟预测框相关的特征图P0 代码 class YoloDetectionBlock(paddle.nn.Layer):def init(self,ch_in,ch_out,is_testTrue):super(YoloDetectionBlock, self).init()assert ch_out % 2 0, \channel {} cannot be divided by 2.format(ch_out)self.conv0 ConvBNLayer(ch_inch_in,ch_outch_out,kernel_size1,stride1,padding0)self.conv1 ConvBNLayer(ch_inch_out,ch_outch_out*2,kernel_size3,stride1,padding1)self.conv2 ConvBNLayer(ch_inch_out*2,ch_outch_out,kernel_size1,stride1,padding0)self.conv3 ConvBNLayer(ch_inch_out,ch_outch_out*2,kernel_size3,stride1,padding1)self.route ConvBNLayer(ch_inch_out*2,ch_outch_out,kernel_size1,stride1,padding0)self.tip ConvBNLayer(ch_inch_out,ch_outch_out*2,kernel_size3,stride1,padding1)def forward(self, inputs):out self.conv0(inputs)out self.conv1(out)out self.conv2(out)out self.conv3(out)route self.route(out)tip self.tip(route)return route, tipNUM_ANCHORS 3 NUM_CLASSES 7 num_filtersNUM_ANCHORS * (NUM_CLASSES 5)backbone DarkNet53_conv_body() detection YoloDetectionBlock(ch_in1024, ch_out512) conv2d_pred paddle.nn.Conv2D(in_channels1024, out_channelsnum_filters, kernel_size1)x np.random.randn(1, 3, 640, 640).astype(float32) x paddle.to_tensor(x) C0, C1, C2 backbone(x) route, tip detection(C0) P0 conv2d_pred(tip)print(P0.shape) 3.6.2 计算预测框是否包含物体的概率 代码 NUM_ANCHORS 3 NUM_CLASSES 7 num_filtersNUM_ANCHORS * (NUM_CLASSES 5)backbone DarkNet53_conv_body() detection YoloDetectionBlock(ch_in1024, ch_out512) conv2d_pred paddle.nn.Conv2D(in_channels1024, out_channelsnum_filters,  kernel_size1)x np.random.randn(1, 3, 640, 640).astype(float32) x paddle.to_tensor(x) C0, C1, C2 backbone(x) route, tip detection(C0) P0 conv2d_pred(tip)reshaped_p0 paddle.reshape(P0, [-1, NUM_ANCHORS, NUM_CLASSES 5, P0.shape[2], P0.shape[3]]) pred_objectness reshaped_p0[:, :, 4, :, :] pred_objectness_probability F.sigmoid(pred_objectness) print(pred_objectness.shape, pred_objectness_probability.shape) 3.6.3 计算预测框位置坐标 代码 NUM_ANCHORS 3 NUM_CLASSES 7 num_filtersNUM_ANCHORS * (NUM_CLASSES 5)backbone DarkNet53_conv_body() detection YoloDetectionBlock(ch_in1024, ch_out512) conv2d_pred  paddle.nn.Conv2D(in_channels1024, out_channelsnum_filters,  kernel_size1)x np.random.randn(1, 3, 640, 640).astype(float32) x paddle.to_tensor(x) C0, C1, C2 backbone(x) route, tip detection(C0) P0 conv2d_pred(tip)reshaped_p0 paddle.reshape(P0, [-1, NUM_ANCHORS, NUM_CLASSES 5, P0.shape[2], P0.shape[3]]) pred_objectness reshaped_p0[:, :, 4, :, :] pred_objectness_probability F.sigmoid(pred_objectness)pred_location reshaped_p0[:, :, 0:4, :, :] print(pred_location.shape) 3.6.4 从P0计算出预测框坐标 代码 def sigmoid(x):return 1./(1.0 np.exp(-x)) def get_yolo_box_xxyy(pred, anchors, num_classes, downsample):batchsize pred.shape[0]num_rows pred.shape[-2]num_cols pred.shape[-1]input_h num_rows * downsampleinput_w num_cols * downsamplenum_anchors len(anchors) // 2pred pred.reshape([-1, num_anchors, 5num_classes, num_rows, num_cols])pred_location pred[:, :, 0:4, :, :]pred_location np.transpose(pred_location, (0,3,4,1,2))anchors_this []for ind in range(num_anchors):anchors_this.append([anchors[ind*2], anchors[ind*21]])anchors_this np.array(anchors_this).astype(float32)pred_box np.zeros(pred_location.shape)for n in range(batchsize):for i in range(num_rows):for j in range(num_cols):for k in range(num_anchors):pred_box[n, i, j, k, 0] jpred_box[n, i, j, k, 1] ipred_box[n, i, j, k, 2] anchors_this[k][0]pred_box[n, i, j, k, 3] anchors_this[k][1]pred_box:, :, :, :, 0 / num_colspred_box:, :, :, :, 1 / num_rowspred_box[:, :, :, :, 2] np.exp(pred_location[:, :, :, :, 2]) * pred_box[:, :, :, :, 2] / input_wpred_box[:, :, :, :, 3] np.exp(pred_location[:, :, :, :, 3]) * pred_box[:, :, :, :, 3] / input_hpred_box[:, :, :, :, 0] pred_box[:, :, :, :, 0] - pred_box[:, :, :, :, 2] / 2.pred_box[:, :, :, :, 1] pred_box[:, :, :, :, 1] - pred_box[:, :, :, :, 3] / 2.pred_box[:, :, :, :, 2] pred_box[:, :, :, :, 0] pred_box[:, :, :, :, 2]pred_box[:, :, :, :, 3] pred_box[:, :, :, :, 1] pred_box[:, :, :, :, 3]pred_box np.clip(pred_box, 0., 1.0)return pred_boxNUM_ANCHORS 3 NUM_CLASSES 7 num_filtersNUM_ANCHORS * (NUM_CLASSES 5)backbone DarkNet53_conv_body() detection YoloDetectionBlock(ch_in1024, ch_out512) conv2d_pred paddle.nn.Conv2D(in_channels1024, out_channelsnum_filters,  kernel_size1)x np.random.randn(1, 3, 640, 640).astype(float32) x paddle.to_tensor(x) C0, C1, C2 backbone(x) route, tip detection(C0) P0 conv2d_pred(tip)reshaped_p0 paddle.reshape(P0, [-1, NUM_ANCHORS, NUM_CLASSES 5, P0.shape[2], P0.shape[3]]) pred_objectness reshaped_p0[:, :, 4, :, :] pred_objectness_probability F.sigmoid(pred_objectness)pred_location reshaped_p0[:, :, 0:4, :, :]anchors [116, 90, 156, 198, 373, 326] pred_boxes get_yolo_box_xxyy(P0.numpy(), anchors, num_classes7, downsample32) print(pred_boxes.shape) 3.6.5 计算物体属于每个类别概率 代码 NUM_ANCHORS 3 NUM_CLASSES 7 num_filtersNUM_ANCHORS * (NUM_CLASSES 5)backbone DarkNet53_conv_body() detection YoloDetectionBlock(ch_in1024, ch_out512) conv2d_pred paddle.nn.Conv2D(in_channels1024, out_channelsnum_filters,  kernel_size1)x np.random.randn(1, 3, 640, 640).astype(float32) x paddle.to_tensor(x) C0, C1, C2 backbone(x) route, tip detection(C0) P0 conv2d_pred(tip)reshaped_p0 paddle.reshape(P0, [-1, NUM_ANCHORS, NUM_CLASSES 5, P0.shape[2], P0.shape[3]]) pred_objectness reshaped_p0[:, :, 4, :, :] pred_objectness_probability F.sigmoid(pred_objectness) pred_location reshaped_p0[:, :, 0:4, :, :] pred_classification reshaped_p0[:, :, 5:5NUM_CLASSES, :, :] pred_classification_probability F.sigmoid(pred_classification) print(pred_classification.shape) 3.7   目标检测实战基于YOLOv3完成林业病虫害检测 3.7.1 得到表示名称字符串和数字类别之间映射关系的字典 代码 INSECT_NAMES  [Boerner, Leconte, Linnaeus, acuminatus, armandi, coleoptera, linnaeus]def get_insect_names():insect_category2id  {}for i, item in enumerate(INSECT_NAMES):insect_category2id[item]  ireturn insect_category2idcname2cid  get_insect_names() cname2cid 3.7.2 读取所有文件标注信息 代码 import os import numpy as np import xml.etree.ElementTree as ETdef get_annotations(cname2cid, datadir):filenames  os.listdir(os.path.join(datadir, annotations, xmls))records  []ct  0for fname in filenames:fid  fname.split(.)[0]fpath  os.path.join(datadir, annotations, xmls, fname)img_file  os.path.join(datadir, images, fid  .jpeg)tree  ET.parse(fpath)if tree.find(id) is None:im_id  np.array([ct])else:im_id  np.array([int(tree.find(id).text)])objs  tree.findall(object)im_w  float(tree.find(size).find(width).text)im_h  float(tree.find(size).find(height).text)gt_bbox  np.zeros((len(objs), 4), dtypenp.float32)gt_class  np.zeros((len(objs), ), dtypenp.int32)is_crowd  np.zeros((len(objs), ), dtypenp.int32)difficult  np.zeros((len(objs), ), dtypenp.int32)for i, obj in enumerate(objs):cname  obj.find(name).textgt_class[i]  cname2cid[cname]_difficult  int(obj.find(difficult).text)x1  float(obj.find(bndbox).find(xmin).text)y1  float(obj.find(bndbox).find(ymin).text)x2  float(obj.find(bndbox).find(xmax).text)y2  float(obj.find(bndbox).find(ymax).text)x1  max(0, x1)y1  max(0, y1)x2  min(im_w - 1, x2)y2  min(im_h - 1, y2)gt_bbox[i]  [(x1x2)/2.0 , (y1y2)/2.0, x2-x11., y2-y11.]is_crowd[i]  0difficult[i]  _difficultvoc_rec  {im_file: img_file,im_id: im_id,h: im_h,w: im_w,is_crowd: is_crowd,gt_class: gt_class,gt_bbox: gt_bbox,gt_poly: [],difficult: difficult}if len(objs) ! 0:records.append(voc_rec)ct  1return recordsTRAINDIR  /home/aistudio/work/insects/train TESTDIR  /home/aistudio/work/insects/test VALIDDIR  /home/aistudio/work/insects/val cname2cid  get_insect_names() records  get_annotations(cname2cid, TRAINDIR)print(records num:{}\n recored[0]:{}.format(len(records), records[0])) 3.7.3 数据预处理 1读取数据 import cv2def get_bbox(gt_bbox, gt_class):MAX_NUM  50gt_bbox2  np.zeros((MAX_NUM, 4))gt_class2  np.zeros((MAX_NUM,))for i in range(len(gt_bbox)):gt_bbox2[i, :]  gt_bbox[i, :]gt_class2[i]  gt_class[i]if i  MAX_NUM:breakreturn gt_bbox2, gt_class2def get_img_data_from_file(record):im_file  record[im_file]h  record[h]w  record[w]is_crowd  record[is_crowd]gt_class  record[gt_class]gt_bbox  record[gt_bbox]difficult  record[difficult]img  cv2.imread(im_file)img  cv2.cvtColor(img, cv2.COLOR_BGR2RGB)assert img.shape[0]  int(h), \image height of {} inconsistent in record({}) and img file({}).format(im_file, h, img.shape[0])assert img.shape[1]  int(w), \image width of {} inconsistent in record({}) and img file({}).format(im_file, w, img.shape[1])gt_boxes, gt_labels  get_bbox(gt_bbox, gt_class)gt_boxes[:, 0]  gt_boxes[:, 0] / float(w)gt_boxes[:, 1]  gt_boxes[:, 1] / float(h)gt_boxes[:, 2]  gt_boxes[:, 2] / float(w)gt_boxes[:, 3]  gt_boxes[:, 3] / float(h)return img, gt_boxes, gt_labels, (h, w) record  records[0] img, gt_boxes, gt_labels, scales  get_img_data_from_file(record) print(img shape:{}, \n gt_labels:{}, \n scales:{}\n.format(img.shape, gt_labels, scales)) (2)数据预处理 a.随机改变暗亮、对比度和颜色 import numpy as np import cv2 from PIL import Image, ImageEnhance import randomdef random_distort(img):def random_brightness(img, lower0.5, upper1.5):e  np.random.uniform(lower, upper)return ImageEnhance.Brightness(img).enhance(e)def random_contrast(img, lower0.5, upper1.5):e  np.random.uniform(lower, upper)return ImageEnhance.Contrast(img).enhance(e)def random_color(img, lower0.5, upper1.5):e  np.random.uniform(lower, upper)return ImageEnhance.Color(img).enhance(e)ops  [random_brightness, random_contrast, random_color]np.random.shuffle(ops)img  Image.fromarray(img)img  ops0img  ops1img  ops2img  np.asarray(img)return imgimport matplotlib.pyplot as plt %matplotlib inline def visualize(srcimg, img_enhance):plt.figure(num2, figsize(6,12))plt.subplot(1,2,1)plt.title(Src Image, color#0000FF)plt.axis(off) plt.imshow(srcimg) srcimg_gtbox  records[0][gt_bbox]srcimg_label  records[0][gt_class]plt.subplot(1,2,2)plt.title(Enhance Image, color#0000FF)plt.axis(off)plt.imshow(img_enhance)image_path  records[0][im_file] print(read image from file {}.format(image_path)) srcimg  Image.open(image_path) srcimg  np.array(srcimg)img_enhance  random_distort(srcimg) visualize(srcimg, img_enhance) b.随机填充 def random_expand(img,gtboxes,max_ratio4.,fillNone,keep_ratioTrue,thresh0.5):if random.random()  thresh:return img, gtboxesif max_ratio  1.0:return img, gtboxesh, w, c  img.shaperatio_x  random.uniform(1, max_ratio)if keep_ratio:ratio_y  ratio_xelse:ratio_y  random.uniform(1, max_ratio)oh  int(h * ratio_y)ow  int(w * ratio_x)off_x  random.randint(0, ow - w)off_y  random.randint(0, oh - h)out_img  np.zeros((oh, ow, c))if fill and len(fill)  c:for i in range©:out_img[:, :, i]  fill[i] * 255.0out_img[off_y:off_y  h, off_x:off_x  w, :]  imggtboxes[:, 0]  ((gtboxes[:, 0] * w)  off_x) / float(ow)gtboxes[:, 1]  ((gtboxes[:, 1] * h)  off_y) / float(oh)gtboxes[:, 2]  gtboxes[:, 2] / ratio_xgtboxes[:, 3]  gtboxes[:, 3] / ratio_yreturn out_img.astype(uint8), gtboxessrcimg_gtbox  records[0][gt_bbox] srcimg_label  records[0][gt_class]img_enhance, new_gtbox  random_expand(srcimg, srcimg_gtbox) visualize(srcimg, img_enhance) c.随机打乱真实框排列顺序 def shuffle_gtbox(gtbox, gtlabel):gt  np.concatenate([gtbox, gtlabel[:, np.newaxis]], axis1)idx  np.arange(gt.shape[0])np.random.shuffle(idx)gt  gt[idx, :]return gt[:, :4], gt[:, 4] 3.7.4 批量数据读取  实验代码 def get_img_size(mode):if (mode  train) or (mode  valid):inds  np.array([0,1,2,3,4,5,6,7,8,9])ii  np.random.choice(inds)img_size  320  ii * 32else:img_size  608return img_sizedef make_array(batch_data):img_array  np.array([item[0] for item in batch_data], dtype  float32)gt_box_array  np.array([item[1] for item in batch_data], dtype  float32)gt_labels_array  np.array([item[2] for item in batch_data], dtype  int32)img_scale  np.array([item[3] for item in batch_data], dtypeint32)return img_array, gt_box_array, gt_labels_array, img_scaleimport paddleclass TrainDataset(paddle.io.Dataset):def  init(self, datadir, modetrain):self.datadir  datadircname2cid  get_insect_names()self.records  get_annotations(cname2cid, datadir)self.img_size  640  #get_img_size(mode)def getitem(self, idx):record  self.records[idx]# print(print: , record)img, gt_bbox, gt_labels, im_shape  get_img_data(record, sizeself.img_size)return img, gt_bbox, gt_labels, np.array(im_shape)def len(self):return len(self.records)train_dataset  TrainDataset(TRAINDIR, modetrain) train_loader  paddle.io.DataLoader(train_dataset, batch_size2, shuffleTrue, num_workers2, drop_lastTrue)img, gt_boxes, gt_labels, im_shape  next(train_loader()) print(img shape:{}\n gt_boxes shape:{}\n gt_labels shape:{}.format(img.shape, gt_boxes.shape, gt_labels.shape)) 3.7.5 模型构建  实验代码 from yolo import YOLOv3 3.7.6 计算损失函数  实验代码 def get_loss(num_classes, outputs, gtbox, gtlabel, gtscoreNone,anchors  [10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326],anchor_masks  [[6, 7, 8], [3, 4, 5], [0, 1, 2]],ignore_thresh0.7,use_label_smoothFalse):losses  []downsample  32for i, out in enumerate(outputs): anchor_mask_i  anchor_masks[i]loss  paddle.vision.ops.yolo_loss(xout,  gt_boxgtbox,  gt_labelgtlabel, gt_scoregtscore,  anchorsanchors,  anchor_maskanchor_mask_i, class_numnum_classes, ignore_threshignore_thresh, downsample_ratiodownsample, use_label_smoothFalse)    losses.append(paddle.mean(loss))  downsample  downsample // 2 return sum(losses)  3.7.7 模型训练  实验代码 import time import os import paddledef get_lr(base_lr  0.0001, lr_decay  0.1):bd  [10000, 20000]lr  [base_lr, base_lr * lr_decay, base_lr * lr_decay * lr_decay]learning_rate  paddle.optimizer.lr.PiecewiseDecay(boundariesbd, valueslr)return learning_rate MAX_EPOCH  1ANCHORS  [10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326]ANCHOR_MASKS  [[6, 7, 8], [3, 4, 5], [0, 1, 2]]IGNORE_THRESH  .7 NUM_CLASSES  7TRAINDIR  /home/aistudio/work/insects/train TESTDIR  /home/aistudio/work/insects/test VALIDDIR  /home/aistudio/work/insects/val paddle.set_device(gpu:0) train_dataset  TrainDataset(TRAINDIR, modetrain) valid_dataset  TrainDataset(VALIDDIR, modevalid) test_dataset  TrainDataset(VALIDDIR, modevalid) train_loader  paddle.io.DataLoader(train_dataset, batch_size10, shuffleTrue, num_workers0, drop_lastTrue, use_shared_memoryFalse) valid_loader  paddle.io.DataLoader(valid_dataset, batch_size10, shuffleFalse, num_workers0, drop_lastFalse, use_shared_memoryFalse) model  YOLOv3(num_classes  NUM_CLASSES)  learning_rate  get_lr() opt  paddle.optimizer.Momentum(learning_ratelearning_rate,momentum0.9,weight_decaypaddle.regularizer.L2Decay(0.0005),parametersmodel.parameters())  # opt  paddle.optimizer.Adam(learning_ratelearning_rate, weight_decaypaddle.regularizer.L2Decay(0.0005), parametersmodel.parameters())if name  main:for epoch in range(MAX_EPOCH):for i, data in enumerate(train_loader()):img, gt_boxes, gt_labels, img_scale  datagt_scores  np.ones(gt_labels.shape).astype(float32)gt_scores  paddle.to_tensor(gt_scores)img  paddle.to_tensor(img)gt_boxes  paddle.to_tensor(gt_boxes)gt_labels  paddle.to_tensor(gt_labels)outputs  model(img)  loss  get_loss(NUM_CLASSES, outputs, gt_boxes, gt_labels, gtscoregt_scores,anchors  ANCHORS,anchor_masks  ANCHOR_MASKS,ignore_threshIGNORE_THRESH,use_label_smoothFalse)       loss.backward()  opt.step() opt.clear_grad()if i % 10  0:timestring  time.strftime(%Y-%m-%d %H:%M:%S,time.localtime(time.time()))print({}[TRAIN]epoch {}, iter {}, output loss: {}.format(timestring, epoch, i, loss.numpy()))if (epoch % 5  0) or (epoch  MAX_EPOCH -1):paddle.save(model.state_dict(), yolo_epoch{}.format(epoch))model.eval()for i, data in enumerate(valid_loader()):img, gt_boxes, gt_labels, img_scale  datagt_scores  np.ones(gt_labels.shape).astype(float32)gt_scores  paddle.to_tensor(gt_scores)img  paddle.to_tensor(img)gt_boxes  paddle.to_tensor(gt_boxes)gt_labels  paddle.to_tensor(gt_labels)outputs  model(img)loss  get_loss(NUM_CLASSES,outputs, gt_boxes, gt_labels, gtscoregt_scores,anchors  ANCHORS,anchor_masks  ANCHOR_MASKS,ignore_threshIGNORE_THRESH,use_label_smoothFalse)if i % 1  0:timestring  time.strftime(%Y-%m-%d %H:%M:%S,time.localtime(time.time()))print({}[VALID]epoch {}, iter {}, output loss: {}.format(timestring, epoch, i, loss.numpy()))model.train() 3.7.8 模型评估函数  实验代码 def box_iou_xyxy(box1, box2):x1min, y1min, x1max, y1max  box1[0], box1[1], box1[2], box1[3]s1  (y1max - y1min  1.) * (x1max - x1min  1.)x2min, y2min, x2max, y2max  box2[0], box2[1], box2[2], box2[3]s2  (y2max - y2min  1.) * (x2max - x2min  1.)xmin  np.maximum(x1min, x2min)ymin  np.maximum(y1min, y2min)xmax  np.minimum(x1max, x2max)ymax  np.minimum(y1max, y2max)inter_h  np.maximum(ymax - ymin  1., 0.)inter_w  np.maximum(xmax - xmin  1., 0.)intersection  inter_h * inter_wunion  s1  s2 - intersectioniou  intersection / unionreturn iou def nms(bboxes, scores, score_thresh, nms_thresh, pre_nms_topk, i0, c0):inds  np.argsort(scores)inds  inds[::-1]keep_inds  []while(len(inds)  0):cur_ind  inds[0]cur_score  scores[cur_ind]# if score of the box is less than score_thresh, just drop itif cur_score  score_thresh:breakkeep  Truefor ind in keep_inds:current_box  bboxes[cur_ind]remain_box  bboxes[ind]iou  box_iou_xyxy(current_box, remain_box)if iou  nms_thresh:keep  Falsebreakif i  0 and c  4 and cur_ind  951:print(suppressed, , keep, i, c, cur_ind, ind, iou)if keep:keep_inds.append(cur_ind)inds  inds[1:]return np.array(keep_inds)def multiclass_nms(bboxes, scores, score_thresh0.01, nms_thresh0.45, pre_nms_topk1000, pos_nms_topk100):batch_size  bboxes.shape[0]class_num  scores.shape[1]rets  []for i in range(batch_size):bboxes_i  bboxes[i]scores_i  scores[i]ret  []for c in range(class_num):scores_i_c  scores_i[c]keep_inds  nms(bboxes_i, scores_i_c, score_thresh, nms_thresh, pre_nms_topk, ii, cc)if len(keep_inds)  1:continuekeep_bboxes  bboxes_i[keep_inds]keep_scores  scores_i_c[keep_inds]keep_results  np.zeros([keep_scores.shape[0], 6])keep_results[:, 0]  ckeep_results[:, 1]  keep_scores[:]keep_results[:, 2:6]  keep_bboxes[:, :]ret.append(keep_results)if len(ret)  1:rets.append(ret)continueret_i  np.concatenate(ret, axis0)scores_i  ret_i[:, 1]if len(scores_i)  pos_nms_topk:inds  np.argsort(scores_i)[::-1]inds  inds[:pos_nms_topk]ret_i  ret_i[inds]rets.append(ret_i)return rets 3.7.9 测试集上模型评估 代码 import json import paddleANCHORS  [10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326]ANCHOR_MASKS  [[6, 7, 8], [3, 4, 5], [0, 1, 2]]VALID_THRESH  0.01NMS_TOPK  400 NMS_POSK  100 NMS_THRESH  0.45NUM_CLASSES  7TESTDIR  /home/aistudio/work/insects/val/images  WEIGHT_FILE  /home/aistudio/yolo_epoch50.pdparams if name  main:model  YOLOv3(num_classesNUM_CLASSES)params_file_path  WEIGHT_FILEmodel_state_dict  paddle.load(params_file_path)model.load_dict(model_state_dict)model.eval()total_results  []test_loader  test_data_loader(TESTDIR, batch_size 1, modetest)for i, data in enumerate(test_loader()):img_name, img_data, img_scale_data  dataimg  paddle.to_tensor(img_data)img_scale  paddle.to_tensor(img_scale_data)outputs  model.forward(img)bboxes, scores  model.get_pred(outputs,im_shapeimg_scale,anchorsANCHORS,anchor_masksANCHOR_MASKS,valid_thresh  VALID_THRESH)bboxes_data  bboxes.numpy()scores_data  scores.numpy()result  multiclass_nms(bboxes_data, scores_data,score_threshVALID_THRESH, nms_threshNMS_THRESH, pre_nms_topkNMS_TOPK, pos_nms_topkNMS_POSK)for j in range(len(result)):result_j  result[j]img_name_j  img_name[j]total_results.append([img_name_j, result_j.tolist()])print(processed {} pictures.format(len(total_results)))print()json.dump(total_results, open(pred_results.json, w)) 3.7.10 模型预测 代码 import numpy as np import matplotlib.pyplot as plt import matplotlib.patches as patches from matplotlib.image import imread import mathINSECT_NAMES  [Boerner, Leconte, Linnaeus, acuminatus, armandi, coleoptera, linnaeus]def draw_rectangle(currentAxis, bbox, edgecolor  k, facecolor  y, fillFalse, linestyle-):rectpatches.Rectangle((bbox[0], bbox[1]), bbox[2]-bbox[0]1, bbox[3]-bbox[1]1, linewidth1,edgecoloredgecolor,facecolorfacecolor,fillfill, linestylelinestyle)currentAxis.add_patch(rect)def draw_results(result, filename, draw_thresh0.5):plt.figure(figsize(10, 10))im  cv2.imread(filename)plt.imshow(im)currentAxisplt.gca()colors  [r, g, b, k, y, c, purple]for item in result:box  item[2:6]label  int(item[0])name  INSECT_NAMES[label]if item[1]  draw_thresh:draw_rectangle(currentAxis, box, edgecolor  colors[label])plt.text(box[0], box[1], name, fontsize12, colorcolors[label])import jsonimport paddleANCHORS  [10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326] ANCHOR_MASKS  [[6, 7, 8], [3, 4, 5], [0, 1, 2]] VALID_THRESH  0.01 NMS_TOPK  400 NMS_POSK  100 NMS_THRESH  0.45NUM_CLASSES  7 if name  main:image_name  /home/aistudio/work/insects/test/images/2642.jpegparams_file_path  /home/aistudio/yolo_epoch50.pdparamsmodel  YOLOv3(num_classesNUM_CLASSES)model_state_dict  paddle.load(params_file_path)model.load_dict(model_state_dict)model.eval()total_results  []test_loader  test_data_loader(image_name, modetest)for i, data in enumerate(test_loader()):img_name, img_data, img_scale_data  dataimg  paddle.to_tensor(img_data)img_scale  paddle.to_tensor(img_scale_data)outputs  model.forward(img)bboxes, scores  model.get_pred(outputs,im_shapeimg_scale,anchorsANCHORS,anchor_masksANCHOR_MASKS,valid_thresh  VALID_THRESH)bboxes_data  bboxes.numpy()scores_data  scores.numpy()results  multiclass_nms(bboxes_data, scores_data,score_threshVALID_THRESH, nms_threshNMS_THRESH, pre_nms_topkNMS_TOPK, pos_nms_topkNMS_POSK)result  results[0] draw_results(result, image_name, draw_thresh0.4) 结果如下 四、总结 YOLOv3采用了全卷积神经网络作为其骨干网络可以在保证网络深度的同时有效缓解梯度消失问题。此外YOLOv3还引入了多尺度特征融合策略将网络划分为三个不同尺度的检测层分别负责检测不同大小的目标。再将不同尺度的特征信息进行融合以提高检测精度。同时YOLOv3还采用了多目标预测的策略即每个预测位置可以同时预测多个目标进一步提高了检测的精度。 因此YOLOv3作为一种高效的目标检测算法通过独特的网络结构和训练策略实现了高效、准确的目标检测。它在速度和精度之间达到了良好的平衡并在实际应用中展现出了广泛的应用前景。