Individual dairy cow identification based on lightweight convolutional neural network

Abstract

In actual farms, individual livestock identification technology relies on large models with slow recognition speeds, which seriously restricts its practical application. In this study, we use deep learning to recognize the features of individual cows. Alexnet is used as a skeleton network for a lightweight convolutional neural network that can recognise individual cows in images with complex backgrounds. The model is improved for multiple multiscale convolutions of Alexnet using the short-circuit connected BasicBlock to fit the desired values and avoid gradient disappearance or explosion. An improved inception module and attention mechanism are added to extract features at multiple scales to enhance the detection of feature points. In experiments, side-view images of 13 cows were collected. The proposed method achieved 97.95% accuracy in cow identification with a single training time of only 6 s, which is one-sixth that of the original Alexnet. To verify the validity of the model, the dataset and experimental parameters were kept constant and compared with the results of Vgg16, Resnet50, Mobilnet V2 and GoogLenet. The proposed model ensured high accuracy while having the smallest parameter size of 6.51 MB, which is 1.3 times less than that of the Mobilnet V2 network, which is famous for its light weight. This method overcomes the defects of traditional methods, which require artificial extraction of features, are often not robust enough, have slow recognition speeds, and require large numbers of parameters in the recognition model. The proposed method works with images with complex backgrounds, making it suitable for actual farming environments. It also provides a reference for the identification of individual cows in images with complex backgrounds.

Lightweight individual cow identification based on Ghost combined with attention mechanism

“…The literature [ 28 ] uses ReXNet 3D for cow behavior recognition and its model size is 14.3 MB with low accuracy, which is 10.69 MB larger than our model, in addition, its FLOPs are 15.8 G, which is also much larger than our model. Finally, compared with our previous work [ 29 ], The model in this study has a good performance improvement in both recognition accuracy and model lightweight, overcoming the deficiencies in recognition accuracy caused by complex backgrounds and speckle distortion from the spatial dimension, improving 0.63 percentage points compared to previous work, and also reducing the model size by 4.97 MB using the linear operation of ghost, making the model more lightweight.…”

Pig Face Recognition Based on Metric Learning by Combining a Residual Network and Attention Mechanism

“…Figure 4b shows the NAM spatial attention submodule, which is calculated as shown in Equation (5). The scaling factor of BN is applied to the spatial dimension to measure the importance of a pixel, which is called pixel normalization.…”

“…However, the traditional methods rely on manually selecting features, and the results of feature extraction directly affect the performance of individual cow recognition. With the development of deep learning and computer vision technology, convolutional neural networks (CNNs) have shown strong feature extraction abilities, and they have been applied to the field of livestock by many researchers [4][5][6][7][8]. Hu et al [9] proposed a novel non-contact cow identification method based on the fusion of deep parts of features by making good use of the YOLO (You Only Look Once) method to detect the cows' heads, trunks, and legs; they built three CNN models to recognize the identities of cows together, and they achieved an accuracy of 98%.…”

Two-stages of segmentation to improve accuracy of deep learning models based on dairy cow morphology

“…A total of 1,020 original images from 102 cows were collected and some examples were shown in Figure 1. It was observed that the dataset scale has a huge impact on the performance of the training network such that when the feature dimensions of the space sample are greater than the training sample's number, the model tends to overfit [19]. Therefore, the training and , 80:20 [21], and 67:33 [22], using python script algorithm in order to determine the optimal model accuracy.…”