A convolutional neural network (CNN) [7] is a deep learning algorithm, used to find patterns present in images. The first few layers of the CNN identify the corners and lines before passing these extracted features down to the neural net. As the CNN gets deeper, it is able to begin recognizing more complex features. Basically, the algorithm performs feature learning and image classification.

Cai et al. [8] presented a back propagation (BP) neural network technique for meter reading recognition. The proposed method is comprised of image enhancement techniques, region of interest extraction, and projection-based character segmentation. The segmented character is passed forward to the back propagation neural network as input and then the extracted feature of each character is matched with the assigned weights and bias values. The network back propagates for better recognition of characters.

Jaiswal et al. [9] claimed improvements in the recognition of the digits using image processing techniques and vertical edge detection algorithms for image segmentation. For the digits learning and recognition phase, they used the neural network.

Gomez et al. [10] designed convolutional neural network architecture to perform segmentation free approach for reading text in real-time images. The proposed method comprised of a seven-layer model having a kernel size of 3 × 3 in all layers and used 2 × 2 kernels in pooling layers for segmentation and text recognition.

Xie et al. [11] used a convolutional neural network-based MD-YOLO (you only look once) model for the detection of a multidirectional vehicle license plate. To achieve better accuracy over other existing methods, they used angle-based rotation and fast intersection over the union prediction evaluation strategy.

Laroca et al. [12] brought improvements in automatic meter reading by using a convolutional neural network. They designed an approach comprised of two steps. Firstly, they used the Fast-YOLO, a convolutional neural network-based object detector model, for the detection of the digit counter. Secondly, they used three different neural network-based approaches (CR-Net, multitask learning, and recurrent neural network) for digit counter recognition.

Ding et al. [13] used deep learning algorithms on various small data sets of daily items for object detection. CNN was used for feature extraction and classification from images and regional proposal networks (RPN) was used for the final object detection. The model parameters were fine-tuned, enabling the deep learning algorithms to perform faster and with better accuracy.

Liu et al. [14] proposed a deep neural model to detect the various objects. They named the model Single Shot Detector (SSD). The model generates several bounding boxes with different sizes and aspect ratios and discretizes the overlapping bounding boxes into default ground truth boxes. For each of the boxes, the model generates the scores for the object present in the box and adjusts with the box having a similar object shape. It also applies the non-maximum suppression technique to discard overlapping boxes and produce the final object detection. Furthermore, the network combines multiple extracted features from images having different resolutions to control various size objects. The proposed multi-box model uses VGG-16 [15] as a base architecture to extract features of autonomous objects deeply in real-time environments.

Ning et al. [16] tried to improve the base SSD model by replacing extra layers of the model with the inception block. The claimed their results had been improved by incorporating the inception block as well as by utilizing batch-normalization and non-maximum suppression techniques for more accurate object detection.

Jin et al. [17] used a vanilla single shot detector model for object detection. They replaced a convolution layer with a max polling layer known as pooling pyramid network (PPN). It is a feature extractor that identifies information from the image and converts the data to a quantifiable format without distorting the image quality. It is an object detection model that is light-weight runs quickly, and performs more accurately when compared to other different single shot detector models.

Tsai [18] used a single shot multi-box detector for digital reading of electricity meters. The pre-trained model worked reasonably well for both collected electricity meter images and the fine-tuned model respectively.

Kim et al. [19] also tried a single shot multi-box detector for the detection and localization of pedestrians and cyclists in the on-road environment. They fine-tuned the model parameters, which increased the overall performance of target object detection and localization compared to the baseline model.

Our first step was to record the target object information that we wanted our model to train on and later detect in the test phase. To do this, we labeled the target object throughout the dataset, using the graphical annotation tool that stores the object information into the XML file. For each image, the XML file contains the position (x; y), width, height, and the class of the object.

Each image with a resolution of 800 × 600 was resized to 300 × 300 pixels to increase the training performance of the model. Then, we applied various image morphological operations to remove noise in the images. The model further converts the XML files into the comma-separated values (CSV) file format. Because the model is unable to process the CSV file format as an input, we needed to convert CSV into a TFRecord file format. This binary file format is easier and faster to process and load during the training phase, so the TFRecord format optimized the data feed. Finally, the label map.pbtxt file is generated which maps every object class label to an integer value and is used in both training and detection phases.

After preprocessing and formatting the image and target object in the dataset, we fed the image to the training model. The training model took the input image and moved it forward to the first convolutional layer, where we resized it in order to feed it forward to the next convolution layers of different filter sizes for feature extraction of the target object as well as, to perform deeper final object detection.

The inception model consists of 42 deep convolutional layers. The computational cost is higher than Google-LeNet [22] and is more efficient than VGG-Net [23]. The model uses different-size convolution kernels that are added to the input image with a maximum pooling layer that decreases computational costs by reducing the number of parameters.

We treated the outputs of the feature extraction with the pooling layer and concatenated them all to feed it forward to the next convolutional layers. This allowed the model to take advantage of multilevel extraction of the features at global 3 × 3 and local 1 × 1 of the target object simultaneously without affecting the network complexity and with improved performance. After concatenation of the features of the target object, the inception model passed these extracted features to the auxiliary network.

The auxiliary network performed object detection by generating more convolutional layers at the end of the first part of the model and decreasing the size of kernels at multiple scales: 10 × 10, 5 × 5, 3 × 3 and 1 × 1. This helps the model predict objects on multiple scales.

The single shot multi-box detector benchmark architecture was further fine-tuned using the Inception model to make the model more robust on meter reading detection. The initial learning rate varies in the range of 0:002–0:004. The batch size value of 12 is used for the sliding window.

We maintained the value of the aspect ratios for use with object localization. Since the input image size required 300 × 300 pixels, the original images were resized from 800 × 600 into 300 × 300. We observed improved accuracy, reported in detail in the results.

The default boxes are matched with ground truth boxes to train the model. Then, each ground truth box is matched with the best-overlapped default box and those bounding boxes are selected for the final object detection whose intersection over union (IOU) is greater than the 0:5 threshold. The model requires a minimum of 100 and a maximum of 500 images for accurate detection of the target object. This training is faster and less data is required. The overall training loss is defined in terms of confidence and localization loss is expressed in equation no (1).

(1) L(x;c;l;g)=1/N(Lconf(x;c)+αLloc(x;l;g))

where (L) denotes the overall loss, (x) represents the input image, (c) stands for confidence, (l) is used for predicted box, and (g) for ground truth boxes. Whereas, (N) denotes the number of default matched boxes, the localization loss (Lloc) depends on smoothing the input image and confidence loss (Lconf) based on the softmax classifier. Typically 80% of the dataset is used for training. The training is performed for 50 epochs with 5,000 steps per epoch. Once the training phase is completed, the proposed model generates checkpoint files and configuration files used during the training. Finally, the inference graph is exported for validation purposes, See Fig. 3.

One hundred (100) images were randomly collected from various sources, the images captured in varying daytime with different camera sensors, viewpoints, and scales.

As already mentioned, we conducted the training of the model on the Google GPU to save the time, but we conducted the tests both on local CPU and Google GPU. For testing the model locally, we downloaded the trained inference graph.pb extension, a binary file that is used for final object detection and label map.pbtxt file that contains the labeling information (x; y), width, height, and class of the target object.

The configuration of the local CPU is Intel Core i3 1.61 GHz processor 8GB RAM and used Python Jupyter notebook on Linux environment whereas, the GPU is Google Colab (Python3 Tesla K80) to examine the robustness of our proposed model in terms of accuracy and speed.

The proposed method processed 12.86 fps to detect the target object on the CPU, whereas the method processed 41.63 fps on the graphical processor (GPU). The baseline methods Mobile-Net processed 7.22 fps on CPU and 25.53 fps on the GPU, the Faster R-CNN processed 0.64 fps on CPU and 8.27 fps on the GPU.

Considering the accuracy, the proposed method detects the target with 96% accuracy, whereas the baseline methods Mobile-Net performs with 92% and the Faster R-CNN with 89% accuracy, outperforming the baseline methods by 4% and 7% respectively, see Tab. 1.

In addition to outperforming the baseline methods in accuracy, the proposed method outperformed both baseline methods with a significant margin both in speed. The proposed method improves about 6 and 13 frames over both the baseline methods when tested on CPU, whereas it shows 16 and 33 frames improvement over GPU. We tested the proposed method over the collected image dataset; the outcome of the method detected the object correctly with significant accuracy, See Fig. 4.

In experiment II, we carried out additional tests to ensure the robustness of the proposed method. We used Tensor board, the web-based precision evaluation tool, to check the precision of the model during the training phase. The board operates by reading Tensor Flow event files generated during the training phase to visualize the performance. The training parameters were linked to the board which updates precision and recall graphs during the training depending upon the number of training iterations.

Using the Tensor board, we were able to visualize the precision of the proposed and the baseline methods. The proposed method precision improves quickly from 800 epochs and reaches to 96% after 4,500 epochs, the baseline Mobile-Net method improves slowly and reaches to 92% after 15,000 epochs, whereas the second baseline method improves slowly and remains stable at 85% for a longer time and finally reaches to 89% after 20,000 epochs, See Fig. 5.

The proposed method for the meter reading requires fewer epochs and present improved results and outperforms the baseline methods.