The post-production stage of agriculture and the food processing industry are progressively being emphasized by advanced technologies. Object counting is an important process in the post-production stage of both sectors. With the progress in computer vision, AI, and deep learning, object counting is becoming more automated, significantly helping in the packaging of food products. This advancement is set to revolutionize how humans interact with object counting, leading to greater accuracy and efficiency in production. Though many deep learning methods on object detection and counting are there, still there is lag in accuracy due to various objects having their own characteristics like shapes, size, color, background, overlapping edges, etc. In this paper, an effective model CountNet is proposed by including a depthwise separable (DS) convolution layer with the traditional convolution neural network (CNN). This layer has two steps and they are depthwise convolution and pointwise convolution. This DS convolution block helps to reduce the computation, to extract complex spatial features and to improve the accuracy. Additionally, we also used squeeze and excitation blocks (SE) to improve the overall performance of the model and to extract the most relevant features for object detection and counting. Our proposed work includes four food products with different sizes and shapes such as almonds, cashews, chocolates, cookies, and apples. Our experimental results show the overall success rate of 98.5%.