Animals and their keeping

Sixty-one camel calves with an age of 12-15 months were used in the current study. These camels were recently transferred from Bulaida Farms, Fujairah, where they were kept in one pen of an area of 900 m² after their weaning, to Al-Marmoum race station, Dubai, United Arab Emirates. These camels were free from any physical abnormalities and many of them showed multiple localized areas of alopecia. Samples of skin scrapping were collected and sent to the Central Veterinary Research Laboratory (CVRL), Dubai, for diagnosis. Ringworm infection on camel skin was diagnosed by a direct microscopic examination according to Almuzaini et al. (2016). Also, samples were diagnosed positively for Ringworm after culturing for 4 weeks [12]. In Al-Marmoum farm, camels were housed in two pens, each covering a total area of 256 square meters. They were fed twice daily on a diet consisting of fresh alfalfa, wheat bran, and flaked barley at rates of 7.0, 0.3, and 0.4 kilograms/head/day, respectively. Additionally, they received a daily oral supplement of 20 grams/head of a mineral and vitamin premix (NeoMix Super®, Neofarma Co., Italy), containing vitamins A, D, E, B1, B2, B6, B12, C, niacin, and pantothenic acid, as well as minerals such as calcium, phosphorus, magnesium, iron, copper, cobalt, iodine, manganese, zinc, and selenium. Access to water and mineral blocks was provided to the camels twice daily.

Experimental setup

This study was done from October to December 2022 and the data was gathered by using four identical cameras, wired full 360-degree monitoring with 1080p Full HD pixels. The flow ability to zoom in or out to make livestock appear closer or farther away during recording was applied. The above cameras were installed around two different areas to monitor camel flow throughout the farm. To ensure the best position of camels in the frames, data collection had to be done in real-time while monitoring camels, utilizing zooming and shooting angle feature Night recordings were omitted from the analysis due to challenges related to low light conditions and issues arising from the peculiar sleeping poses of camels, making them unsuitable for meaningful analysis.

All the camels were visually divided into two categories: Infected and non-infected cases. The infected cases were subdivided according to the degree of infection into (A) mild, in which a total number of < 50 infection spots were detected on an area of the camel body (neck, hump, abdomen, rump, or legs) ; (B) moderate, in which a total number of 50-100 infection spots were detected on 1-2 body areas; (C) severe, in which a total number of >100 infection spots were detected on more than 2 areas of the camel body.

Data Collection

Videos capturing camel activity from all four cameras were carefully processed to extract specific frames showcasing the intricate details of the camels' skin. Subsequently, these selected frames were saved as high-quality images. A total number of 941 images for healthy and non-healthy camels were analyzed.  These images were manually filtered down to 801 depending on the resolution, angle, and position of the camel, before being split 70%/18%/12% for use in training, validation, and testing, respectively. The next step involves labeling these images to create a labeled dataset for training purposes. The dataset of 801 images was manually annotated and afterward as shown in Table 1.

Image Segmentation and Yolov7 Model

Images were annotated using the polygon tool in Roboflow [13] which provides more granular information to the training pipeline about the data. This data were used to train a model and, in turn, resulted in a more accurate and effective model. A data processing flowchart (Fig. 1) illustrates the steps taken to annotate images and generate annotated files, which were subsequently exported in a suitable file format. Furthermore, the workflow diagram for the input data and the result data is shown in Fig 2. Image segmentation is a difficult problem in computer vision for a variety of reasons, including the complexity of the images themselves, their texture, and their homogeneity [14]. Before training the models, all photos underwent pre-processing. To fit the input size of the YOLOv7 architecture, all images were scaled to 640×640 pixels. To ensure consistent input across all images, the pixel values are standardized to the range (0, 1).

The YOLOv7 architecture was trained using PyTorch. The training process involved 50 epochs with a batch size of 16 [15]. Adam optimizer was utilized with a learning rate of 0.001 and a weight decay of 0.0005. The mean squared error (MSE) loss function was employed during training. Following the training phase, the performance of the YOLOv7 model was assessed first on the validation set, comprising 144 images, and then on the test set, consisting of 97 images. Further, the evaluation has been done. Pressure (P), recall (R), and mean average precision (mAP) were all important benchmarks for gauging the model’s accuracy. As demonstrated in Equation (1), the number of accurate positive predictions (TP) was a measure of precision. The equation was employed to calculate a model’s recall, also known as sensitivity, representing the ratio of true positives to the total number of predicted positives in a dataset [2]. Equation (3) expressed the harmonic mean of accuracy and memory recall. The F1 score, when contrasted with recall and precision, presented several advantages. For extremely low values of precision and recall, F1 normalized the measurements. The Python programming language and the Pytorch framework were employed. Table 2 shows the hyperparameters that were used during the training process to develop the model.

(1) Precision: It represents the ratio of true positives (TP) to the sum of true positives and false positives (FP) in predictions.

(2) Recall: It signifies the ratio of true positives (TP) to the sum of true positives and false negatives (FN) in actual positives.

(3) F1-score: This is the harmonic mean of precision and recall, calculated as F1 = 2 * (precision * recall) / (precision + recall).

Results

The infected camels were 36 out of a total of 61 camels with an incidence of 59%. These infected camels were subdivided into 6, 5, and 25 for mild, moderate, and severe, respectively (Table 3). The algorithm achieved an average precision (AP) of 0.944 on the validation set, with effectiveness in distinguishing between normal and infected skin. The training loss graph shows the pattern of training loss over the 50 epochs of training. The training loss started at around 0.0099. The final training loss was approximately 0.0078. Similarly, the validation loss graph shows the pattern of validation loss over the 50 epochs of training. The validation loss started at around 0.00772 and decreased steadily throughout training. During the training, the loss in training was noticed to be 0.0491. The final validation loss was approximately 0.0075 (Figure 3). The confusion matrix shows the classification results of the YOLOv7 algorithm on the validation set. The matrix shows that the algorithm correctly classified infected and non-infected skin images of camels. Also, it could distinguish the degree of infection. The model's visual detection and segmentation results on the testing dataset were analyzed to assess its performance, as shown in Figure 3. The precision-recall curve illustrates the YOLOv7 algorithm's performance regarding precision and recall. The curve shows that the algorithm achieved average precision and recall of 0.94325, and 0.99526 respectively (Table 4).

Figures & Tables

Discussion