主要算法：目标检测网络RetinaNet

本方案需要完成前置模型转换工作采取的方案为：pt—onnx—tflite（tflite为了完成部署到移动端）

完成转换后将模型部署至aidlux平台，完成实时视频检测。部分代码如下： `def process_img(img, target_size=640, max_size=2000, multiple=32, keep_ratio=True, NCHW=True, ToTensor=True): img = img[128:512, 0:480] img = cv2.resize(img, (640, 512), interpolation=cv2.INTER_LINEAR) im_shape = img.shape im_size_min = np.min(im_shape[0:2]) im_size_max = np.max(im_shape[0:2])

resize with keep_ratio

if keep_ratio: im_scale = float(target_size) / float(im_size_min) if np.round(im_scale * im_size_max) > max_size: im_scale = float(max_size) / float(im_size_max) im_scale_x = np.floor(img.shape[1] * im_scale / multiple) * multiple / img.shape[1] im_scale_y = np.floor(img.shape[0] * im_scale / multiple) * multiple / img.shape[0] image_resized = cv2.resize(img, None, None, fx=im_scale_x, fy=im_scale_y, interpolation=cv2.INTER_LINEAR) im_scales = np.array([im_scale_x, im_scale_y, im_scale_x, im_scale_y]) im = image_resized / 255.0 # np.float64 im = im.astype(np.float32) PIXEL_MEANS =(0.485, 0.456, 0.406) # RGB format mean and variances PIXEL_STDS = (0.229, 0.224, 0.225) im -= np.array(PIXEL_MEANS) im /= np.array(PIXEL_STDS) im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB) # BGR2RGB if NCHW: im = np.transpose(im, (2, 0, 1)).astype(np.float32) # [SAI-KEY] TensorFlow use input with NHWC. im = im[np.newaxis, ...] if ToTensor: im = torch.from_numpy(im) return im, im_scales else: return None `

此代码用于对输入的图片进行预处理，使其能够被神经网络处理。具体来说，输入的图片首先被裁剪和缩放到指定大小，然后通过减去 RGB 归一化值的均值和标准差进行归一化，接着将 RGB 通道转为 BGR 通道，最后改变图像的输入格式，将其转为 NCHW 格式的张量（N 代表 Batch size，C 代表通道数，H 代表高度，W 代表宽度）并转为 PyTorch 的 Tensor 类型。如果 keep_ratio=True，则图像的宽高比被保持不变。返回处理后的图像以及缩放比例。如果 keep_ratio=False，则返回 None。

if name=="main": tflite_model = '/home/R-RetinaNet/models/r-retinanet.tflite' # 定义输入输出shape in_shape = [1 * 640 * 800 * 3 * 4] # HWC, float32 out_shape = [1 * 53325 * 8 * 4] # 8400: total cells, 52 = 48(num_classes) + 4(xywh), float32

# AidLite初始化
aidlite = aidlite_gpu.aidlite()
# 加载R-RetinaNet模型
res = aidlite.ANNModel(tflite_model, in_shape, out_shape, 4, -1) # Infer on -1: cpu, 0: gpu, 1: mixed, 2: dsp
# print(res)
'''
读取手机实时摄像头数据
'''
cap = cvs.VideoCapture(0)
frame_id = 0
while True:
    frame = cap.read()
    if frame is None:
        continue
    im, im_scales = process_img(frame, NCHW=False, ToTensor=False)  # im: NHWC
    frame_id += 1
    if frame_id % 3 != 0:
        continue



    aidlite.setInput_Float32(im, 800, 640)
    # 推理
    aidlite.invoke()
    preds = aidlite.getOutput_Float32(0)
    preds = preds.reshape(1, 8, (int)(preds.shape[0]/8))
    # 后解算
    output = np.transpose(preds, (0, 2, 1))

    # 创建Anchor
    im_anchor = np.transpose(im, (0, 3, 1, 2)).astype(np.float32)
    anchors_list = []
    anchor_generator = Anchors(ratios = np.array([0.2, 0.5, 1, 2, 5]))
    original_anchors = anchor_generator(im_anchor)   # (bs, num_all_achors, 5)
    anchors_list.append(original_anchors)

    # 解算输出
    decode_output = decoder(im_anchor, anchors_list[-1], output[..., 5:8], output[..., 0:5], thresh=0.2, nms_thresh=0.2, test_conf=None)

    # 重构输出
    scores = decode_output[0].reshape(-1, 1)
    classes = decode_output[1].reshape(-1, 1)
    boxes = decode_output[2]
    boxes[:, :4] = boxes[:, :4] / im_scales
    if boxes.shape[1] > 5:   
        boxes[:, 5:9] = boxes[:, 5:9] / im_scales
    dets = np.concatenate([classes, scores, boxes], axis=1)

    # 过滤类别
    keep = np.where(classes > 0)[0]
    dets =  dets[keep, :]

    # 转换坐标('xywha'->'xyxyxyxy')
    res = sort_corners(rbox_2_quad(dets[:, 2:]))

    # cv绘图.
    for k in range(dets.shape[0]):
        cv2.line(frame, (int(res[k, 0]), int(res[k, 1])), (int(res[k, 2]), int(res[k, 3])), (0, 255, 0), 3)
        cv2.line(frame, (int(res[k, 2]), int(res[k, 3])), (int(res[k, 4]), int(res[k, 5])), (0, 255, 0), 3)
        cv2.line(frame, (int(res[k, 4]), int(res[k, 5])), (int(res[k, 6]), int(res[k, 7])), (0, 255, 0), 3)
        cv2.line(frame, (int(res[k, 6]), int(res[k, 7])), (int(res[k, 0]), int(res[k, 1])), (0, 255, 0), 3)

    cvs.imshow(frame)

此代码实现了通过手机摄像头实时检测图像中的文本行，首先初始化 AidLite，并加载 R-RetinaNet 模型。然后进入摄像头读取和处理的循环中，先调用 process_img 对图像进行预处理，然后将预处理后的图像输入给模型进行推理，再根据模型输出进行解析、过滤和转换坐标，最后在原图上绘制文本行框并显示出来。其中使用了 opencv 绘制框和显示图片。效果如下：