Estimation of the Gender Ratio of Chickens Based on Computer Vision: Dataset and Exploration

Abstract

<jats:p>The gender ratio of free-range chickens is considered as a major animal welfare problem in commercial broiler farming. Free-range chicken producers need to identify chicken gender to estimate the economic value of their flock. However, it is challenging for farmers to estimate the gender ratio of chickens efficiently and accurately, since the environmental background is complicated and the chicken number is dynamic. Moreover, manual estimation is likely double counts or missed count and thus is inaccurate and time consuming. Hence, automated methods that can lead to results efficiently and accurately replace the identification abilities of a chicken gender expert, working in a farm environment, are beneficial to the industry. The contributions in this paper include: (1) Building the world’s first chicken gender classification database annotated manually, which comprises 800 chicken flock images captured on a farm and 1000 single chicken images separated from the flock images by an object detection network, labelled with gender information. (2) Training a rooster and hen classifier using a deep neural network and cross entropy in information theory to achieve an average accuracy of 96.85%. The evaluation of the algorithm performance indicates that the proposed automated method is practical for the gender classification of chickens on the farm environment and provides a feasible way of thinking for the estimation of the gender ratio.</jats:p>

Practices and Applications of Convolutional Neural Network-Based Computer Vision Systems in Animal Farming: A Review

“…To correct the distortion and uniform objects in an image, some standard references (e.g., straight line, marker, and checkerboard) can be annotated in images [ 79 , 107 , 124 , 125 ], and then different transformation models (e.g., isometry, similarity, affine, and projective) can be applied to make the whole images match the respective references [ 123 ]. In addition, some restoration models can be applied to remove the noises processed during data acquisition, encoding, and transmission and to improve image quality [ 126 ]. Median filtering that averages intensities based on neighboring pixels was used in general [ 52 , 110 , 122 ], and other models that included two-dimensional gamma function-based image adaptive correction algorithm and several layers of neural networks were also available for the denoising [ 63 , 126 , 127 ].…”

“…In addition, some restoration models can be applied to remove the noises processed during data acquisition, encoding, and transmission and to improve image quality [ 126 ]. Median filtering that averages intensities based on neighboring pixels was used in general [ 52 , 110 , 122 ], and other models that included two-dimensional gamma function-based image adaptive correction algorithm and several layers of neural networks were also available for the denoising [ 63 , 126 , 127 ]. Some images contain natural noises that are hard to be denoised.…”

“…Some images contain natural noises that are hard to be denoised. Adding known noises (with predefined mean and variance) including Gaussian, salt, and pepper noises may improve the denoising efficiency [ 63 , 126 ]. As for depth sensing, uneven lighting conditions and object surfaces, dirt, or dust may result in holes within an object in an image [ 115 , 128 ].…”

“…The combinations had three categories. The first one was to combine multiple CNNs, such as YOLO + AlexNet [ 151 ], YOLO V2 + ResNet50 [ 98 ], and YOLO V3 + (AlexNet, VGG16, VGG19, ResNet18, ResNet34, DenseNet121) [ 126 ]; the second one was to combine CNNs with regular machine learning models, such as fully connected network (FCN) + support vector machine (SVM) [ 68 ], Mask region-based CNN (mask R-CNN) + kernel extreme learning machine (KELM) [ 90 ], Tiny YOLO V2 + SVM [ 121 ], VGG16 + SVM [ 49 ], YOLO + AlexNet + SVM [ 74 ], and YOLO V3 + [SVM, K-nearest neighbor (KNN), decision tree classifier (DTC)] [ 152 ]; and the third one was to combine CNNs with other deep learning techniques, such as [Convolutional 3 dimension (C3D), VGG16, ResNet50, DenseNet169, EfficientNet] + long short-term memory (LSTM) [ 84 ], Inception V3 + bidirectional LSTM (BiLSTM) [ 41 ], and Inception V3 + LSTM [ 153 ], YOLO V3 + LSTM [ 152 ]. In all these combinations, CNN typically played roles of feature extractors in the first stage, and then other models utilized these features to make classifications.…”

“…These include losses for classification, detection, and segmentation [ 51 , 111 ]. Accuracy over different steps is also applied during training [ 126 ]. Although evaluation metrics during training can help optimize models, final performance of models during testing is of major concern.…”