Applications of deep learning in visual recognition

Abstract

Animal welfare research has raised concerns regarding the intensification of farm animal housing systems that offer limited opportunity for movement. However, no currently available automated tracking software is able to efficiently and accurately track dairy cow movement across stall-based housing systems. Applying deep learning models to location tracking provides an opportunity for accurate and timely measurement of cow movement within the housing environment. The objective of this study was to develop and validate a location tracking tool to monitor the movement of dairy cows in their tie-stalls using a deep learning approach. Twenty-four lactating Holstein cows were video recorded for a continuous 24-h period on weeks 1, 2, 3, 6, 8, and 10. Individual images showing the in-stall position of each cow were extracted from each 24-h recording at a rate of one image per minute. Three coordinates on each cow were manually annotated on the image sequences to track the location of the left hip, the right hip, and the neck. The final dataset used to validate the deep learning approach consisted of 199,100 Red-Green-Blue images with manual coordinate annotations. The dataset was separated into training and validation sets. Variants of the following deep learning models were tested: VGG Net, Resnet, GoogLeNet, and DenseNet. Model performance was expressed in terms of pixel error for each coordinate annotated from the validation image set. Pixel error was converted to a standard measure in cm using the average pix/cm ratio for each cow in each week. ResNet18 with augmented labels significantly outperformed all other models tested. For the validation image set, the average error from all 3 coordinates was equivalent to a 0.74 cm error in actual physical placement of the coordinates within the stall environment. Based on this high degree of accuracy, the model may be used to analyze the activity patterns of individual cows for optimization of stall spaces and improved ease of movement. Synthetic Aperture Radar (SAR) imagery captures the physical properties of the Earth by transmitting microwave signals to its surface and analyzing the backscattered signal. It does not depends on sunlight and therefore can be obtained in any condition, such as nighttime and cloudy weather. However, SAR images are noisier than light images and so far it is not clear the level of performance that a modern recognition system could achieve. This work presents an analysis of the performance of deep learning models for the task of land segmentation using SAR images. We present segmentation results on the task of classifying four different land categories (urban, water, vegetation and farm) on six Canadian sites (Montreal, Ottawa, Quebec, Saskatoon, Toronto and Vancouver), with three state-of-the-art deep learning segmentation models. Results show that when enough data and variety on the land appearance are available, deep learning models can achieve an excellent performance despite the high input noise.

Date	22 Mar 2023
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Patrick Cardinal (Supervisor), Marco Pedersoli (Co-supervisor) & Elsa Vasseur (Co-supervisor)