利用AidLux实现热成像电力巡检项目作业展示及流程介绍

---

前言

对于刚入门的计算机视觉算法工程师而言，目标检测算法的数据集准备+训练+部署依然存在着不小的难度，而Aidlux的出现很好的解决了这一个问题。Aidlux可以将我们的安卓设备进一步改造成边缘计算设备。本文为参加aidlux训练营的学习和作业展示，用以记录该项目的整体流程，如有不当之处，请在评论区联系作者。项目完整代码和数据集均由aidlux提供，不知道放上是否合适，故只放上核心代码，完整资源可去aidlux开发者社区官网寻找。

本项目为基于Aidlux+r-retinanet+tflite，在安卓手机小米6上实现热成像电力训练项目。具体内容为：通过r-retinanet对绝缘子等电力设施进行旋转目标检测。

---

一、模型训练及onnx模型导出

博主没有训练r-retinanet模型的条件，因此使用的是训练营老师提供的onnx模型（感谢老师),通过下方代码将其导出为能在安卓设备上运行的tflite模型，核心代码如下。

def onnx_converter(onnx_model_path:str,  output_path:str=None, 
                    input_node_names:list=None, output_node_names:list=None,
                    need_simplify:bool=True, target_formats:list = ['keras', 'tflite'],
                    native_groupconv:bool=False,
                    weight_quant:bool=False, int8_model:bool=False, image_root:str=None,
                    int8_mean:list or float = [123.675, 116.28, 103.53], int8_std:list or float = [58.395, 57.12, 57.375])->float:
    if not isinstance(target_formats, list) and  'keras' not in target_formats and 'tflite' not in target_formats:
        raise KeyError("'keras' or 'tflite' should in list")
    
    model_proto = load_onnx_modelproto(onnx_model_path, input_node_names, output_node_names, need_simplify)

    keras_model = keras_builder(model_proto, native_groupconv)

    if 'tflite' in target_formats:
        tflite_model = tflite_builder(keras_model, weight_quant, int8_model, image_root, int8_mean, int8_std)

    onnx_path, model_name = os.path.split(onnx_model_path)
    if output_path is None:
        output_path = onnx_path
    output_path = os.path.join(output_path, model_name.split('.')[0])

    keras_model_path = None
    if 'keras' in target_formats:
        keras_model_path = output_path + ".h5"
        keras_model.save(keras_model_path)
        LOG.info(f"keras model saved in {keras_model_path}")

    tflite_model_path = None
    if 'tflite' in target_formats:
        tflite_model_path = output_path + ".tflite"
        with open(tflite_model_path, "wb") as fp:
            fp.write(tflite_model)

    convert_result = {"keras":keras_model_path, "tflite":tflite_model_path, "keras_error":0, "tflite_error":0}
    # ignore quantization model
    if int8_model:
        return convert_result
    
    error_dict = {}
    try:
        error_dict = get_elements_error(model_proto, keras_model_path, tflite_model_path)
        keras_error, tflite_error = error_dict.get("keras", None), error_dict.get("tflite", None)
        if keras_error:
            if keras_error > 1e-2:
                LOG.error("h5 model elements' max error has reached {:^.4E}, but convert is done, please check {} carefully!".format(keras_error, keras_model_path))
            elif keras_error > 1e-4:
                LOG.warning("h5 model elements' max error is {:^.4E}, pass, h5 saved in {}".format(keras_error, keras_model_path))
            else:
                LOG.info("h5 model elements' max error is {:^.4E}, pass, h5 saved in {}".format(keras_error, keras_model_path))
        if tflite_error:
            if tflite_error > 1e-2:
                LOG.error("tflite model elements' max error has reached {:^.4E}, but convert is done, please check {} carefully!".format(tflite_error, tflite_model_path))
            elif tflite_error > 1e-4:
                LOG.warning("tflite model elements' max error is {:^.4E}, pass, tflite saved in {}".format(tflite_error, tflite_model_path))
            else:
                LOG.info("tflite model elements' max error is {:^.4E}, pass, tflite saved in {}".format(tflite_error, tflite_model_path))
    except:
        LOG.warning("convert is successed, but model running is failed, please check carefully!")
    
    convert_result["keras_error"] = error_dict.get("keras", None)
    convert_result["tflite_error"] = error_dict.get("tflite", None)
    return convert_result

import os
import sys
sys.path.append("/2T/001_AI/1003_YOLOv8/005_Misc/onnx2tflite-main")
from converter import onnx_converter

import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
 
def onnx2tflite(onnx_path):
    onnx_converter(
        onnx_model_path = onnx_path,
        need_simplify = False,
        output_path = os.path.dirname(onnx_path), 
        target_formats = ['tflite'], # or ['keras'], ['keras', 'tflite']
        weight_quant = False,
        int8_model = False,
        int8_mean = None,
        int8_std = None,
        image_root = None
        )

if __name__ == "__main__":
    onnx2tflite("./r-retinanet.onnx")

导出成功后，得到tflite模型。

二、vscode远程连接安卓设备进行aidlux开发

手机上需要先下载AidLux App，进入后点击Cloud_ip图标,会出现自己手机的aidlux ip,注意我们通过vscode远程连接时，端口号不是8000，而是9022。

1.remote-shh安装

在左侧扩展中搜索remote-shh，安装成功后左侧会出现小电脑图标.

按照顺序进行config配置

配置内容如下，HostName填上上面的IP地址，随后输入密码即可连接成功，将电脑上的代码传到手机里即可进行开发。

2.调用手机后置摄像头进行检测

aidlux是通过cvs包捕获摄像头图像的。

if __name__=="__main__":
        
    ''' 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
    # out_shape = [1 * 55425 * 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 '''
    tflite_model = '/home/AidLux_Deploy/models/r-retinanet.tflite'
    res = aidlite.ANNModel(tflite_model, in_shape, out_shape, 4, -1) # Infer on -1: cpu, 0: gpu, 1: mixed, 2: dsp

    print(res)
    '''
    读取手机后置摄像头
    '''
    cap = cvs.VideoCapture(0)
    frame_id = 0
    while True:
        frame = cap.read()

        if frame is None:
            continue
        frame_id += 1
        if frame_id % 30 != 0:
            continue
        time0 = time.time()
        im, im_scales = process_img(frame, NCHW=False, ToTensor=False)  # im: NHWC
        ''' 设定输入输出 '''
        aidlite.setInput_Float32(im, 800, 640)
        ''' 启动推理 '''
        aidlite.invoke()
        ''' 捕获输出 '''
        preds = aidlite.getOutput_Float32(0)
        post_process_img(preds,frame,im)
        # print('ending')

3.过程中遇到的旋转框偏移问题

模型的输入大小是固定的800x640，博主一开始通过老师提供的代码直接运行时报错，debug时发现我的小米6经过预处理图像的尺寸变为852x640.然后想着直接resize到固定大小，应该就没问题了，结果检测发现结果明显存在偏移,如左图。

博主发现是检测结果的post_process_img模块，将预测结果对应回原图时用到了之前预处理的缩放因子,因此直接resize会出现这种状况，正好想到之前再跑yolov8算法时，通过letterbox方法是能正确显示目标框的，于是对代码中的尺寸处理和解码模块进行了替换，最终显示结果正常，如下图。

process模块代码修改如下

def letterbox(img, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleFill=False, scaleup=True, stride=32):
    # Resize and pad image while meeting stride-multiple constraints
    shape = img.shape[:2]  # current shape [height, width]
    if isinstance(new_shape, int):
        new_shape = (new_shape, new_shape)

    # Scale ratio (new / old)
    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
    if not scaleup:  # only scale down, do not scale up (for better test mAP)
        r = min(r, 1.0)

    # Compute padding
    ratio = r, r  # width, height ratios
    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
    if auto:  # minimum rectangle
        dw, dh = np.mod(dw, stride), np.mod(dh, stride)  # wh padding
    elif scaleFill:  # stretch
        dw, dh = 0.0, 0.0
        new_unpad = (new_shape[1], new_shape[0])
        ratio = new_shape[1] / shape[1], new_shape[0] / shape[0]  # width, height ratios

    dw /= 2  # divide padding into 2 sides
    dh /= 2

    if shape[::-1] != new_unpad:  # resize
        img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_LINEAR)
    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
    img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
    return img, ratio, (dw, dh)
 resized_img, ratio ,_= letterbox(img, (640, 800), stride=None, auto=False)

post_process_img模块代码修改如下

scale_boxes方法如下

def scale_boxes(img1_shape, boxes, img0_shape, ratio_pad=None):
    if ratio_pad is None:  # calculate from img0_shape
        gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1])  # gain  = old / new
        pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2  # wh padding
    else:
        gain = ratio_pad[0][0]
        pad = ratio_pad[1]

    boxes[..., [0, 2]] -= pad[0]  # x padding
    boxes[..., [1, 3]] -= pad[1]  # y padding
    boxes[..., :4] /= gain
    # clip_boxes(boxes, img0_shape)
    return boxes

---

总结

以上是参加Aidlux电力巡检训练营的记录，对整体流程初步了解，核心代码和模型都是老师提供，还需要进行深入研究，才能体会到项目中真正存在的坑。